A new system can automatically retouch images in the style of a professional photographer. It can run on a cellphone and display retouched images in real-time. Courtesy of the researchers (edited by MIT News)

The data captured by today’s digital cameras is often treated as the raw material of a final image. Before uploading pictures to social networking sites, even casual cellphone photographers might spend a minute or two balancing color and tuning contrast, with one of the many popular image-processing programs now available.

This week at Siggraph, the premier digital graphics conference, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory and Google are presenting a new system that can automatically retouch images in the style of a professional photographer. It’s so energy-efficient, however, that it can run on a cellphone, and it’s so fast that it can display retouched images in real-time, so that the photographer can see the final version of the image while still framing the shot.

The same system can also speed up existing image-processing algorithms. In tests involving a new Google algorithm for producing high-dynamic-range images, which capture subtleties of color lost in standard digital images, the new system produced results that were visually indistinguishable from those of the algorithm in about one-tenth the time — again, fast enough for real-time display.

The system is a machine-learning system, meaning that it learns to perform tasks by analyzing training data; in this case, for each new task it learned, it was trained on thousands of pairs of images, raw and retouched.

The work builds on an earlier project from the MIT researchers, in which a cellphone would send a low-resolution version of an image to a web server. The server would send back a “transform recipe” that could be used to retouch the high-resolution version of the image on the phone, reducing bandwidth consumption.

“Google heard about the work I’d done on the transform recipe,” says Michaël Gharbi, an MIT graduate student in electrical engineering and computer science and first author on both papers. “They themselves did a follow-up on that, so we met and merged the two approaches. The idea was to do everything we were doing before but, instead of having to process everything on the cloud, to learn it. And the first goal of learning it was to speed it up.”

Short cuts

In the new work, the bulk of the image processing is performed on a low-resolution image, which drastically reduces time and energy consumption. But this introduces a new difficulty, because the color values of the individual pixels in the high-res image have to be inferred from the much coarser output of the machine-learning system.

In the past, researchers have attempted to use machine learning to learn how to “upsample” a low-res image, or increase its resolution by guessing the values of the omitted pixels. During training, the input to the system is a low-res image, and the output is a high-res image. But this doesn’t work well in practice; the low-res image just leaves out too much data.

Gharbi and his colleagues — MIT professor of electrical engineering and computer science Frédo Durand and Jiawen Chen, Jon Barron, and Sam Hasinoff of Google — address this problem with two clever tricks. The first is that the output of their machine-learning system is not an image; rather, it’s a set of simple formulae for modifying the colors of image pixels. During training, the performance of the system is judged according to how well the output formulae, when applied to the original image, approximate the retouched version.

Taking bearings

The second trick is a technique for determining how to apply those formulae to individual pixels in the high-res image. The output of the researchers’ system is a three-dimensional grid, 16 by 16 by 8. The 16-by-16 faces of the grid correspond to pixel locations in the source image; the eight layers stacked on top of them correspond to different pixel intensities. Each cell of the grid contains formulae that determine modifications of the color values of the source images.

That means that each cell of one of the grid’s 16-by-16 faces has to stand in for thousands of pixels in the high-res image. But suppose that each set of formulae corresponds to a single location at the center of its cell. Then any given high-res pixel falls within a square defined by four sets of formulae.

Roughly speaking, the modification of that pixel’s color value is a combination of the formulae at the square’s corners, weighted according to distance. A similar weighting occurs in the third dimension of the grid, the one corresponding to pixel intensity.

The researchers trained their system on a data set created by Durand’s group and Adobe Systems, the creators of Photoshop. The data set includes 5,000 images, each retouched by five different photographers. They also trained their system on thousands of pairs of images produced by the application of particular image-processing algorithms, such as the one for creating high-dynamic-range (HDR) images. The software for performing each modification takes up about as much space in memory as a single digital photo, so in principle, a cellphone could be equipped to process images in a range of styles.

Finally, the researchers compared their system’s performance to that of a machine-learning system that processed images at full resolution rather than low resolution. During processing, the full-res version needed about 12 gigabytes of memory to execute its operations; the researchers’ version needed about 100 megabytes, or one-hundredth as much. The full-resolution version of the HDR system took about 10 times as long to produce an image as the original algorithm, or 100 times as long as the researchers’ system.

“This technology has the potential to be very useful for real-time image enhancement on mobile platforms,” says Barron. “Using machine learning for computational photography is an exciting prospect but is limited by the severe computational and power constraints of mobile phones. This paper may provide us with a way to sidestep these issues and produce new, compelling, real-time photographic experiences without draining your battery or giving you a laggy viewfinder experience.”

In episode four of season three Neil introduces us to the ideas behind the bias variance dilemma (and how how we can think about it in our daily lives). Plus, we answer a listener question about how to make sure your neural networks don’t get fooled. Our guest for this episode is Jeff Dean,  Google Senior Fellow in the Research Group, where he leads the Google Brain project. We talk about a closet full of robot arms (the arm farm!), image recognition for diabetic retinopathy, and equality in data and the community.

Fun Fact: Geoff Hinton’s distant relative invented the word tesseract. (How cool is that. Seriously.)

• Talking Machines: Overfitting and asking ecological questions, with Tom Dietterich
• Talking Machines: Restricted Boltzmann Machines, with Eric Lander
• Talking Machines: Automatic Translation and t-SNE, with Hal Daume
• Talking Machines: Generative art and Hamiltonian Monte Carlo, with Doug Eck
• Talking Machines: Machine learning and the Flint water crisis, with Jake Abernethy
• Talking Machines: Gaussian processes and OpenAI, with IIya Sutskever

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

Image: MIT News

Laparoscopy is a surgical technique in which a fiber-optic camera is inserted into a patient’s abdominal cavity to provide a video feed that guides the surgeon through a minimally invasive procedure. Laparoscopic surgeries can take hours, and the video generated by the camera — the laparoscope — is often recorded. Those recordings contain a wealth of information that could be useful for training both medical providers and computer systems that would aid with surgery, but because reviewing them is so time consuming, they mostly sit idle.

Researchers at MIT and Massachusetts General Hospital hope to change that, with a new system that can efficiently search through hundreds of hours of video for events and visual features that correspond to a few training examples.

In work they presented at the International Conference on Robotics and Automation this month, the researchers trained their system to recognize different stages of an operation, such as biopsy, tissue removal, stapling, and wound cleansing.

But the system could be applied to any analytical question that doctors deem worthwhile. It could, for instance, be trained to predict when particular medical instruments — such as additional staple cartridges — should be prepared for the surgeon’s use, or it could sound an alert if a surgeon encounters rare, aberrant anatomy.

“Surgeons are thrilled by all the features that our work enables,” says Daniela Rus, an Andrew and Erna Viterbi Professor of Electrical Engineering and Computer Science and senior author on the paper. “They are thrilled to have the surgical tapes automatically segmented and indexed, because now those tapes can be used for training. If we want to learn about phase two of a surgery, we know exactly where to go to look for that segment. We don’t have to watch every minute before that. The other thing that is extraordinarily exciting to the surgeons is that in the future, we should be able to monitor the progression of the operation in real-time.”

Joining Rus on the paper are first author Mikhail Volkov, who was a postdoc in Rus’ group when the work was done and is now a quantitative analyst at SMBC Nikko Securities in Tokyo; Guy Rosman, another postdoc in Rus’ group; and Daniel Hashimoto and Ozanan Meireles of Massachusetts General Hospital (MGH).

Representative frames

The new paper builds on previous work from Rus’ group on “coresets,” or subsets of much larger data sets that preserve their salient statistical characteristics. In the past, Rus’ group has used coresets to perform tasks such as deducing the topics of Wikipedia articles or recording the routes traversed by GPS-connected cars.

In this case, the coreset consists of a couple hundred or so short segments of video — just a few frames each. Each segment is selected because it offers a good approximation of the dozens or even hundreds of frames surrounding it. The coreset thus winnows a video file down to only about one-tenth its initial size, while still preserving most of its vital information.

For this research, MGH surgeons identified seven distinct stages in a procedure for removing part of the stomach, and the researchers tagged the beginnings of each stage in eight laparoscopic videos. Those videos were used to train a machine-learning system, which was in turn applied to the coresets of four laparoscopic videos it hadn’t previously seen. For each short video snippet in the coresets, the system was able to assign it to the correct stage of surgery with 93 percent accuracy.

“We wanted to see how this system works for relatively small training sets,” Rosman explains. “If you’re in a specific hospital, and you’re interested in a specific surgery type, or even more important, a specific variant of a surgery — all the surgeries where this or that happened — you may not have a lot of examples.”

Selection criteria

The general procedure that the researchers used to extract the coresets is one they’ve previously described, but coreset selection always hinges on specific properties of the data it’s being applied to. The data included in the coreset — here, frames of video — must approximate the data being left out, and the degree of approximation is measured differently for different types of data.

Machine learning can be thought of as a problem of approximation, however. In this case, the system had to learn to identify similarities between frames of video in separate laparoscopic feeds that denoted the same phases of a surgical procedure. The metric of similarity that it arrived at also served to assess the similarity of video frames that were included in the coreset, to those that were omitted.

“Interventional medicine — surgery in particular — really comes down to human performance in many ways,” says Gregory Hager, a professor of computer science at Johns Hopkins University who investigates medical applications of computer and robotic technologies. “As in many other areas of human endeavor, like sports, the quality of the human performance determines the quality of the outcome that you achieve, but we don’t know a lot about, if you will, the analytics of what creates a good surgeon. Work like what Daniela is doing and our work really goes to the question of: Can we start to quantify what the process in surgery is, and then within that process, can we develop measures where we can relate human performance to the quality of care that a patient receives?”

“Right now, efficiency” — of the kind provided by coresets — “is probably not that important, because we’re dealing with small numbers of these things,” Hager adds. “But you could imagine that, if you started to record every surgery that’s performed — we’re talking tens of millions of procedures in the U.S. alone — now it starts to be interesting to think about efficiency.”

Credit: Wikipedia Commons

In episode three, season three of Talking Machines, we dive into overfitting, take a listener question about unbalanced data and speak with Professor (Emeritus) Tom Dietterich from Oregon State University.

If you enjoyed this episode, you may also want to listen to:

• Talking Machines: Restricted Boltzmann Machines, with Eric Lander
• Talking Machines: Automatic Translation and t-SNE, with Hal Daume
• Talking Machines: Generative art and Hamiltonian Monte Carlo, with Doug Eck
• Talking Machines: Machine learning and the Flint water crisis, with Jake Abernethy
• Talking Machines: Gaussian processes and OpenAI, with IIya Sutskever

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

A GelSight sensor attached to a robot’s gripper enables the robot to determine precisely where it has grasped a small screwdriver, removing it from and inserting it back into a slot, even when the gripper screens the screwdriver from the robot’s camera. Photo: Robot Locomotion Group at MIT

Eight years ago, Ted Adelson’s research group at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) unveiled a new sensor technology, called GelSight, that uses physical contact with an object to provide a remarkably detailed 3-D map of its surface.

Now, by mounting GelSight sensors on the grippers of robotic arms, two MIT teams have given robots greater sensitivity and dexterity. The researchers presented their work in two papers at the International Conference on Robotics and Automation last week.

In one paper, Adelson’s group uses the data from the GelSight sensor to enable a robot to judge the hardness of surfaces it touches — a crucial ability if household robots are to handle everyday objects.

In the other, Russ Tedrake’s Robot Locomotion Group at CSAIL uses GelSight sensors to enable a robot to manipulate smaller objects than was previously possible.

The GelSight sensor is, in some ways, a low-tech solution to a difficult problem. It consists of a block of transparent rubber — the “gel” of its name — one face of which is coated with metallic paint. When the paint-coated face is pressed against an object, it conforms to the object’s shape.

The metallic paint makes the object’s surface reflective, so its geometry becomes much easier for computer vision algorithms to infer. Mounted on the sensor opposite the paint-coated face of the rubber block are three colored lights and a single camera.

“[The system] has colored lights at different angles, and then it has this reflective material, and by looking at the colors, the computer … can figure out the 3-D shape of what that thing is,” explains Adelson, the John and Dorothy Wilson Professor of Vision Science in the Department of Brain and Cognitive Sciences.

In both sets of experiments, a GelSight sensor was mounted on one side of a robotic gripper, a device somewhat like the head of a pincer, but with flat gripping surfaces rather than pointed tips.

Contact points

For an autonomous robot, gauging objects’ softness or hardness is essential to deciding not only where and how hard to grasp them but how they will behave when moved, stacked, or laid on different surfaces. Tactile sensing could also aid robots in distinguishing objects that look similar.

In previous work, robots have attempted to assess objects’ hardness by laying them on a flat surface and gently poking them to see how much they give. But this is not the chief way in which humans gauge hardness. Rather, our judgments seem to be based on the degree to which the contact area between the object and our fingers changes as we press on it. Softer objects tend to flatten more, increasing the contact area.

The MIT researchers adopted the same approach. Wenzhen Yuan, a graduate student in mechanical engineering and first author on the paper from Adelson’s group, used confectionary molds to create 400 groups of silicone objects, with 16 objects per group. In each group, the objects had the same shapes but different degrees of hardness, which Yuan measured using a standard industrial scale.

Then she pressed a GelSight sensor against each object manually and recorded how the contact pattern changed over time, essentially producing a short movie for each object. To both standardize the data format and keep the size of the data manageable, she extracted five frames from each movie, evenly spaced in time, which described the deformation of the object that was pressed.

Finally, she fed the data to a neural network, which automatically looked for correlations between changes in contact patterns and hardness measurements. The resulting system takes frames of video as inputs and produces hardness scores with very high accuracy. Yuan also conducted a series of informal experiments in which human subjects palpated fruits and vegetables and ranked them according to hardness. In every instance, the GelSight-equipped robot arrived at the same rankings.

Yuan is joined on the paper by her two thesis advisors, Adelson and Mandayam Srinivasan, a senior research scientist in the Department of Mechanical Engineering; Chenzhuo Zhu, an undergraduate from Tsinghua University who visited Adelson’s group last summer; and Andrew Owens, who did his PhD in electrical engineering and computer science at MIT and is now a postdoc at the University of California at Berkeley.

Obstructed views

The paper from the Robot Locomotion Group was born of the group’s experience with the Defense Advanced Research Projects Agency’s Robotics Challenge (DRC), in which academic and industry teams competed to develop control systems that would guide a humanoid robot through a series of tasks related to a hypothetical emergency.

Typically, an autonomous robot will use some kind of computer vision system to guide its manipulation of objects in its environment. Such systems can provide very reliable information about an object’s location — until the robot picks the object up. Especially if the object is small, much of it will be occluded by the robot’s gripper, making location estimation much harder. Thus, at exactly the point at which the robot needs to know the object’s location precisely, its estimate becomes unreliable. This was the problem the MIT team faced during the DRC, when their robot had to pick up and turn on a power drill.

“You can see in our video for the DRC that we spend two or three minutes turning on the drill,” says Greg Izatt, a graduate student in electrical engineering and computer science and first author on the new paper. “It would be so much nicer if we had a live-updating, accurate estimate of where that drill was and where our hands were relative to it.”

That’s why the Robot Locomotion Group turned to GelSight. Izatt and his co-authors — Tedrake, the Toyota Professor of Electrical Engineering and Computer Science, Aeronautics and Astronautics, and Mechanical Engineering; Adelson; and Geronimo Mirano, another graduate student in Tedrake’s group — designed control algorithms that use a computer vision system to guide the robot’s gripper toward a tool and then turn location estimation over to a GelSight sensor once the robot has the tool in hand.

In general, the challenge with such an approach is reconciling the data produced by a vision system with data produced by a tactile sensor. But GelSight is itself camera-based, so its data output is much easier to integrate with visual data than the data from other tactile sensors.

In Izatt’s experiments, a robot with a GelSight-equipped gripper had to grasp a small screwdriver, remove it from a holster, and return it. Of course, the data from the GelSight sensor don’t describe the whole screwdriver, just a small patch of it. But Izatt found that, as long as the vision system’s estimate of the screwdriver’s initial position was accurate to within a few centimeters, his algorithms could deduce which part of the screwdriver the GelSight sensor was touching and thus determine the screwdriver’s position in the robot’s hand.

“I think that the GelSight technology, as well as other high-bandwidth tactile sensors, will make a big impact in robotics,” says Sergey Levine, an assistant professor of electrical engineering and computer science at the University of California at Berkeley. “For humans, our sense of touch is one of the key enabling factors for our amazing manual dexterity. Current robots lack this type of dexterity and are limited in their ability to react to surface features when manipulating objects. If you imagine fumbling for a light switch in the dark, extracting an object from your pocket, or any of the other numerous things that you can do without even thinking — these all rely on touch sensing.”

“Software is finally catching up with the capabilities of our sensors,” Levine adds. “Machine learning algorithms inspired by innovations in deep learning and computer vision can process the rich sensory data from sensors such as the GelSight to deduce object properties. In the future, we will see these kinds of learning methods incorporated into end-to-end trained manipulation skills, which will make our robots more dexterous and capable, and maybe help us understand something about our own sense of touch and motor control.”

In episode two of season three Neil takes us through the basics on dropout, we chat about the definition of inference (It’s more about context than you think!) and hear an interview with Jennifer Chayes of Microsoft.

If you liked this article, you may also want to read:

• Talking Machines: Restricted Boltzmann Machines, with Eric Lander
• Talking Machines: Automatic Translation and t-SNE, with Hal Daume
• Talking Machines: Generative art and Hamiltonian Monte Carlo, with Doug Eck
• Talking Machines: Machine learning and the Flint water crisis, with Jake Abernethy
• Talking Machines: Gaussian processes and OpenAI, with IIya Sutskever

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

In this article, we explain in plain language machine learning.

Learning to learn

In 1959, Arthur Samuel, a pioneer in the field of machine learning (ML) defined it as the “field of study that gives computers the ability to learn without being explicitly programmed”.

ML can be understood as computational methods that use experience to improve performance or to make accurate predictions. In this case, experience refers to past information or data that is available to us, which has been labelled and categorized. As with any computational exercise, the quality and amount of the data will be crucial to the accuracy of the predictions that will be made.

Looking through this lens, ML seems to be a lot like statistical modelling. In statistical modelling, we collect data, verify that it is clean — in other words, that we have completed, corrected, or deleted any incomplete, incorrect, or irrelevant parts of the data — and then use this clean dataset to test hypotheses and make predictions and forecasts. The idea behind statistical modelling is the attempt to represent complex issues in relatively generalizable terms, which is to say, terms that explain most events studied. Effectively, we programme the algorithm to perform certain functions based on the data we submit. Put differently, the algorithm is static. It needs a programmer to tell it what to do when it is fed with data. So far, it makes indeed sense to use this approach when activated by the programmer.

But with ML, the procedure is flipped. Rather than preselecting a model and feeding data into it, in ML it is the data that determines which analytic technique should be selected to best perform the task at hand. In other words, the computer uses the data that it has to select and train the algorithm. Hence the algorithm is no longer static. It analyses the data to which it is exposed, makes a determination on the best course of action, and then acts. In essence, it “learns” from the data and in doing so, knowledge can be extracted from the data.

This method of learning is based on repetition. Remember that an algorithm is nothing more than a set of instructions that a computer uses to transform an input into a particular output. Thus in ML, the learning aspect is just an algorithm repeating its execution operation over and over again and making slight adjustments until a certain set of conditions is met. The litmus test of a learning algorithm is when it is able to predict when new data is given to it on which it had not previously been trained.

Evolution of ML

Data obviously plays a primary role in this methodological process. More importantly, it is the structure of the data that determines how the learning process will occur. It is here that we see the three-levels of ML:

Supervised ML

Here the computer is trained on data that is well labelled. This means that the data is already tagged with the correct label or the correct outcome. For example, if we were to teach a computer to distinguish between the picture of a cat and a dog, then we would tag the data in the following way:

This labelling process should be done by the programmer and having learned the difference, the ML algorithm can now classify new information that is given to it and determine if the new image is that of a dog or a cat.

Based on this simplistic method, supervised ML can be used to perform much more complicated operations. One use is understanding how to read figures and alphabets. The way one person writes the number “1” or the letter “A” will not be the same way as another person does.

By feeding the computer with vast amounts of labelled examples of the number “1” or “A”, we can train the algorithm to see the various versions these figures. The computer thus begins to learn the variations and becomes increasingly competent at understanding these patterns.

Today, computers are better than humans in recognizing such patterns of handwriting. The larger the dataset, the better trained the algorithm. Once trained, the algorithm is given new data and uses its past experience to predict an outcome.

Unsupervised ML

This is where the algorithm is trained using a dataset that does not have any labels. The algorithm is not told what the data represents. In this case, the learning process is dependent on the identification of patterns that are repeatedly created in the data. Using the cat and dog example, the algorithm begins to separate the images it receives based on the inherent characteristics of dogs.

In unsupervised learning, the algorithms must use methods of estimation based on inferential statistics to discover patterns, relationships and correlations with the raw, unlabelled dataset. As patterns are identified, the algorithm uses statistics to identify boundaries within the dataset. Data with similar patterns are grouped together, creating subsets of data. As the classification process continues, the algorith begins to understand the dataset it is analysing, allowing it to predict the categorization of future data.

This clustering of data can automate decision making, adding a layer of sophistication to unsupervised learning. More importantly, it allows us to leverage data in a new way. What we lack in knowledge we make up for in data.

Reinforcement ML

Reinforcement learning is like unsupervised ML in that the training data is also unlabelled. However, when asked a question about the data, the outcome is graded – so there is still a level of supervision. The algorithm is presented with data that lacks labels, but is given an example with a positive or negative result. This positive or negative grade provides a feedback loop to the algorithm allowing it to determine if the solution it is providing is solving a problem or not. Effectively, it is the computerised version of human trial and error learning.

Reinforcement ML is often used to make strategies. As decisions lead to consequences, the output action is prescriptive, and not just descriptive, as in unsupervised learning. This kind of learning has been used to train computers how to play games. This is the idea behind the company DeepMind, acquired by Google in 2014, which trained which they trained their algorithm to learn how to play Atari.

It went on to create AlphaGo which defeated the best human professional Go player 4-1 in the game of Go, one of the most complex games in the world.

Implications for businesses

Today machine learning is being used in a number of areas. Google’s self-driving car was developed using machine learning and today machines can lip read faster than humans. ML has been infiltrating almost every sector of finance in recent years. For instance, ML is being used for algorithmic trading, analysing time series data, portfolio management, fraud detection, customer service, news analysis, investment strategy construction, etc.

But the real power of machine learning is unleashed with neural networks.

This article was originally on The Conversation. Click here to access the original.

Imagine how easy it would be to learn skating, if only it doesn’t hurt everytime you fall. Unfortunately, we, humans,  don’t have that option. Robots, however, can now “learn” their skills on a simulation platform without being afraid of crashing into a wall. Yes, “it learns“! This is possible with the reinforcement learning algorithms provided by OpenAI Gym and the ROS Development Studio.

You can now train your robot to navigate through an environment filled with obstacles just based on the sensor inputs, with the help of OpenAI Gym. In April 2016, OpenAI introduced “Gym”, a platform for developing and comparing reinforcement learning algorithms. Reinforcement learning is an area of machine learning that allows an intelligent agent (for example, robot) to learn the best behaviors in an environment by trial-and-error. The agent takes actions in an environment so as to maximize its rewards. We have deployed the gym_gazebo package from Erle Robotics. in the ROS Development Studio. It enables the users to test their reinforcement learning for their robots in Gazebo.

How to Train your Robot

Robot training to navigate through an environment with obstacles

In this example we will be seeing how a turtlebot is able to learn navigating through an environment without hitting an obstacle. The turtlebot will use a reinforcement learning method known as Q-learning.

There are four environments already available for the user to test their simulations with. These environments can be launched using the respective launch files:

• GazeboCircuitTurtlebotLidar_v0.launch
• GazeboCircuit2TurtlebotLidar_v0.launch
• GazeboRoundTurtlebotLidar_v0.launch
• GazeboMazeTurtlebotLidar_v0.launch

The images of the different environments are given below:

The various environments already available:

Circuit	Circuit2
Round	Maze

The user is requested to try out the existing environment before developing their own environments for training the robot. An environment is where the robot’s possible actions and rewards are defined. For example, in the available environments, there are three possible actions for the Turtlebot robot:

– Forward ( with a reward of 5 points)
– Left ( with a reward of 1 point)
– Right ( with a reward of 1 point)

If it collides into the walls, then the training episode ends (with a penalty of 200 points). The turtlebot has to learn to navigate through the environment, based on the rewards obtained from different episodes. This can be achieved using the Q-learning algorithm. Let us see, how it can be done using the ROS Development Studio.

Using the ROS Development Studio for training the robot

First, we have to set the path in the jupyter-notebook as given below:

import sys
sys.path.append("/usr/local/lib/python2.7/dist-packages/")
sys.path.append("/home/ubuntu/gym-gazebo")
sys.path.append("/home/user/catkin_ws/src/gym_construct/src")
%matplotlib inline


The python scripts in the folder gym_construct/src/ help us simulate the reinforcement learning techniques for a Turtlebot. Currently, the number of episodes has been set to 20.
Feel free to increase the number of episodes in the python scripts (usually up to 5000) to actually train the robot to navigate the environment completely.Run the  only the script corresponding to the environment:

Run the python corresponding to the environment:

## Circuit-2 Environment --> Q-learning
%run /home/user/catkin_ws/src/gym_construct/src/circuit2_turtlebot_lidar_qlearn.py


The robot undergoes several training episodes and each of these episodes are rewarded based on the number of steps taken by the robot before hitting the environment. We will be able to see that the robot incrementally increases its rewards over time compared to its previous versions. With a very large number of episodes, the robot will learn to navigate the environment without hitting the obstacle.

Plotting the learning curve of the robot

The machine learning algorithm generates the output files in the output directory specified in the python script file. In order to plot the curve, we run the display_plot.py. But before that, don’t forget to restart the kernel and set the path once again.

As the number of episodes increases, we will see that the robot’s mean rewards also increases. The user can choose his own robot, environments, action and rewards for testing his reinforcement learning algorithms in OpenAI Gym and RDS. Watch the video below for more on this.

Video describing the procedure to run OpenAI Gym on RDS

I hope you were able to follow the tutorial.  So, that’s all folks. It’s now up to you to develop and test reinforcement learning algorithms in OpenAI Gym and RDS.

Have fun training your robot!

Last week I had the pleasure of debating the question “does AI pose a threat to society?” with friends and colleagues Christian List, Maja Pantic and Samantha Payne. The event was organised by the British Academy and brilliantly chaired by the Royal Society’s director of science policy Claire Craig. Here follows my opening statement:

My opening statement

One Friday afternoon in 2009 I was called by a science journalist at, I recall, the Sunday Times. He asked me if I knew that there was to be a meeting of the AAAI to discuss robot ethics. I said no I don’t know of this meeting. He then asked “are you surprised they are meeting to discuss robot ethics” and my answer was no. We talked some more and agreed it was actually a rather dull story: a case of scientists behaving responsibly. I really didn’t expect the story to appear but checked the Sunday paper anyway, and there in the science section was the headline Scientists fear revolt of killer robots. (I then spent the next couple of days on the radio explaining that no, scientists do not fear a revolt of killer robots.)

So, fears of future super intelligence – robots taking over the world – are greatly exaggerated: the threat of an out-of-control super intelligence is a fantasy – interesting for a pub conversation perhaps. It’s true we should be careful and innovate responsibly, but that’s equally true for any new area of science and technology. The benefits of robotics and AI are so significant, the potential so great, that we should be optimistic rather than fearful. Of course robots and intelligent systems must be engineered to very high standards of safety for exactly the same reasons that we need our washing machines, cars and airplanes to be safe. If robots are not safe people will not trust them. To reach it’s full potential what robotics and AI needs is a dose of good old fashioned (and rather dull) safety engineering.

In 2011 I was invited to join a British Standards Institute working group on robot ethics, which drafted a new standard BS 8611 Guide to the ethical design of robots and robotic systems, published in April 2016. I believe this to be the world’s first standard on ethical robots.

Also in 2016 the very well regarded IEEE standards association – the same organization that gave us WiFi – launched a Global initiative on Ethical Considerations in AI and Autonomous Systems. The purpose of this Initiative is to ensure every technologist is educated and empowered to prioritize ethical considerations in the design and development of autonomous and intelligent systems; in a nutshell, to ensure ethics are baked in. In December we published Ethically Aligned Design: A Vision for Prioritizing Human Well Being with AI and Autonomous Systems. Within that initiative I’m also leading a new standard on transparency in autonomous systems, based on the simple principle that it should always be possible to find out why an AI or robot made a particular decision.

We need to agree ethical principles, because they are needed to underpin standards – ways of assessing and mitigating the ethical risks of robotics and AI. But standards needs teeth and in turn underpin regulation. Why do we need regulation? Think of passenger airplanes; the reason we trust them is because it’s a highly regulated industry with an amazing safety record, and robust, transparent processes of air accident investigation when things do go wrong. Take one example of a robot that we read a lot about in the news – the Driverless Car. I think there’s a strong case for a driverless car equivalent of the CAA, with a driverless car accident investigation branch. Without this it’s hard to see how driverless car technology will win public trust.

Does AI pose a threat to society? No. But we do need to worry about the down to earth questions of present day rather unintelligent AIs; the ones that are deciding our loan applications, piloting our driverless cars or controlling our central heating. Are those AIs respecting our rights, freedoms and privacy? Are they safe? When AIs make bad decisions, can we find out why? And I worry too about the wider societal and economic impacts of AI. I worry about jobs of course, but actually I think there is a bigger question: how can we ensure that the wealth created by robotics and AI is shared by all in society?

Thank you.

The featured image of this article was used to advertise the BA’s series of events on the theme Robotics, AI and Society. The reason I draw your attention to it here is that one of the many interesting questions to the panel was about the way that AI tends to be visualised in the media. This kind of human face coalescing (or perhaps emerging) from the atomic parts of the AI seems to have become a trope for AI. Is it a helpful visualisation of the human face of AI, or does it mislead to an impression that AI has human characteristics?

Credit: nutonomy

Over the past few weeks, we’ve blogged about how machine learning is transforming research, nuclear decommissioning, and astronomy. Building on our “Ask the Experts” panel discussion on driverless cars last July, how might machine learning shape the way we move through and interact with the immediate environment around us?

As the last in our series of blog posts on machine learning in research, we spoke to Dr Nathan Griffiths to find out more about machine learning in transport. Nathan is a Reader in the Department of Computer Science at the University of Warwick, whose research into the application of machine learning for autonomous vehicles (or “driverless cars”) has been supported by a Royal Society University Research Fellowship.

What inspired you to study science? And why Computer Science particularly?

I have always been interested in how things work and understanding the patterns we see in the world around us. Science helps us to understand patterns, provides insights into how things work, and ultimately enables us to understand how complex structures can arise from simple rules. Computer Science appeals because it can be both a creative and an analytical science – we can use computers to create patterns, or use computers as tools to understand our world, including the patterns that exist within it, and predict patterns which may arise.

How is machine learning used in your research?

My research uses machine learning to give insights into how objects or people interact and how patterns emerge and evolve. This involves the use of computational models, known as agent-based modelling, to simulate the actions and interactions of entities (such as drivers or autonomous vehicles) to see how their decisions affect the wider systems in which they operate.

More specifically, machine learning can be used to analyse how cooperation emerges through interactions between objects. In the context of autonomous vehicles, there are many situations in which it is desirable to establish a cooperative environment, without the need for centralised control. For example, who should give way at a road junction where the rules are ambiguous or typically misinterpreted, such as flashing headlights to let another driver proceed? In this case, machine learning can be trained to predict what decisions a driver or autonomous vehicle will make. Machine learning algorithms will examine previous behaviours and learn from these behaviours, to then predict what will happen in the future. Many of the initial predictions will be incorrect, however the algorithm will modify its next predictions based on what it has learnt and continue to reduce its mistakes until it can make accurate predictions. An accurate algorithm could then be used to inform the decisions vehicles make and predict vehicle journeys and routes.

Modern vehicles are equipped with numerous sensors to collect vast amounts of data during vehicle journeys. For example cameras provide live images to the vehicle which are fed into an object-recognition machine learning algorithm which learns to identify and distinguish between entities such as pedestrians, cyclists or other vehicles on the road. The images can tell the vehicle how close the objects are and change the vehicle’s behaviour accordingly, such as by reducing speed to stay a safe distance from the car in front. The use of these data also enables transport to be made more efficient, such as through automatic traffic light signal detection and live notifications of parking opportunities which could help to reduce congestion.

What are the challenges in using machine learning in your field of research?

The techniques for extracting information from data and building predictive models are advanced, but perhaps the greatest challenge is to support the interface between machine learning and human users, which is likely to lead to greater public confidence in something like autonomous vehicles.

I also expect to see advances in techniques for dealing with “concept drift”, where changes in the world   need to be accompanied by changes in a machine learnt model. Concept drift is the phenomena whereby predictions may become less accurate over time if the models do not adapt to changes in the environment. This could be an issue if learning models are not able to adapt to changes sufficiently quickly to react.

Currently, autonomous vehicles are good at detecting the world around them and reacting as necessary to changes on the road. However the next challenge is to enable machine learning systems to not only react to changes and adapt to previously unseen environments (such as a swerving car or the vehicle being driven in adverse conditions) but be able to understand the probabilities of what it thinks might happen. For example, a car which is driving alongside a busy footpath will not just react to a pedestrian stepping out into the road, but will understand that there is a chance that this might happen and actively take measures to reduce risk in case it does, perhaps by positioning itself further away from the curb.

So what are the future prospects for machine learning?

The future is very bright for machine learning. Machine learning is fundamental to the development of driverless vehicles and is rapidly transforming the automotive world; vehicle manufacturers are very keen to get computer scientists on board for the betterment of their projects.

Machine learning is already central to how our vehicles work, underpinning systems such as lane-keeping assistance, reactive cruise control and automated vehicle personalisation and recognition. There is a wide and increasing range of applications that machine learning will allow us to tackle and it is enabling us to develop fully autonomous vehicles, which will be ubiquitous in the future. I also envisage that autonomous vehicles will be connected, such that the information and insights that one vehicle has learnt can be transferred to another; which may prove particularly useful for managing concept drift.

The Royal Society is undertaking a project about machine learning and its impact for the UK economy and society. To find out more please visit our project page.

This post was originally published on the Royal Society. Click here to view the original post.

Talking Machines is entering its third season and going through some changes. Our founding host Ryan is moving on and in his place, Neil Lawrence of Amazon is taking over as co-host. We say thank you and goodbye to Ryan with an interview about his work.

If you liked this article, you may also want to read:

• Talking Machines: Restricted Boltzmann Machines, with Eric Lander
• Talking Machines: Automatic Translation and t-SNE, with Hal Daume
• Talking Machines: Generative art and Hamiltonian Monte Carlo, with Doug Eck
• Talking Machines: Machine learning and the Flint water crisis, with Jake Abernethy
• Talking Machines: Gaussian processes and OpenAI, with IIya Sutskever

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

Heavy Water Components Test Reactor decommissioning.

How can we create robots that can carry out important tasks in dangerous environments? Machine learning is supporting advances in the field of robotics. To find out more, we talked to Dr Rustam Stolkin, Royal Society Industry Fellow for Nuclear Robotics, Professor of Robotics at the University of Birmingham, and Director at A.R.M Robotics Ltd, about his work combining machine learning and robotics to create practical solutions to nuclear problems.

What drew you to engineering and robotics?

There are many definitions of engineering, but the one I like is “the creation of artefacts for the benefit of mankind”. Engineering is a way of using science to be creative and to create novel technologies which can bring major societal and economic benefit.

Dr Rustam Stolkin. Source: University of Birmingham/YouTube

Robotics, in particular, attracted me because robots are inherently inter-disciplinary: they use mechanics and electronics, as well as computational elements, such as machine learning and artificial intelligence. This has allowed me to become an “expert generalist”; I specialise in being extremely cross-disciplinary. At the University of Birmingham, I have research collaborations spanning eight different schools across three colleges on the campus.

Tell us about your research?

My work focuses on developing advanced robotics technologies for nuclear decommissioning, demolishing legacy nuclear sites, and safely packaging, storing and monitoring any radiological or contaminated materials. The UK contains 4.9 million tonnes of legacy nuclear waste, mainly by-products of weapons production during the Cold War, alongside a smaller contribution from nuclear power use for civilian electricity generation. Cleaning up the UK’s legacy waste is the biggest environmental remediation project across Europe; it could take over 100 years and cost up to £220billion.

Factories are highly structured environments, where robots can be pre-programmed to make repetitive motions. This has supported the revolution in robotics that we’ve seen since the 1980s. In contrast, there has been surprisingly little use of robots in the nuclear industry, aside from a very small number of robotic vehicles deployed after the Fukushima disaster, and a very few cases of deploying bespoke robotic manipulators, controlled in rudimentary ways by joystick teleoperation.

This is in part because legacy nuclear sites, some of them over 60 years old, are highly complex and unknown environments, such as contaminated buildings that have been sealed off for decades. They have an enormous diversity of scenes, materials, and objects that nuclear decommissioning robots must interact with in complex ways, such as by grasping and cutting objects.

We have only recently discovered how to use state-of-the-art machine learning methods to help robots learn adaptive and transferable skills which they can apply in new environments. This allows them to perceive and handle new kinds of objects which they have not seen before. For example, using machine learning, a robot can be trained how to grasp and pick up one object, and then use this knowledge to adapt the grasp for a new object with a different shape.

Recent “deep learning” approaches might offer part of the solution in navigating these complex environments. However such learning approaches can rely on large amounts of “training data” – labelled examples of images of objects for example – which can be used to train the robot so that it can recognise similar objects in future. It is still very unclear how to generate sufficient training data in order to enable robots to learn how to behave in useful ways in complex real environments.

Why do we need robots to carry out nuclear decommissioning?

Contaminated materials are currently decommissioned by human workers in air-fed plastic suits, which is not only unsustainable but can be dangerous. Using robots means humans do not need to enter highly hazardous work environments, and means that zones that humans cannot enter (even in protective suits) due to high gamma radiation can still be decommissioned. Using robots also reduces “secondary nuclear waste”; for every one container filled with actual primary nuclear waste, more than ten containers become filled with contaminated plastic suits, respirators, rubber gloves, and other “secondary waste” from human entries.

How have you used machine learning to develop robots capable of nuclear decommissioning?

My lab has used robots to perform fully autonomous grasping and manipulation of nuclear waste-like objects and grasping of moving objects.

We also use state-of-the-art deep-learning neural networks to do real-time 3D reconstruction of a room, and simultaneously recognise and label materials such as concrete, metal, wood and fabric. This helps us develop ways for robots to characterise nuclear environments and nuclear waste objects in real-time. We are extending these methods to also combine visual camera information with other kinds of data from radiation, thermal and other sensors.

Where do you see machine learning going within your field of research?

Robotics engineers often argue that machine learning has only been used on relatively simple problems so far. Achieving machine learning of real robotic actions, with real objects in real scenes, is a significantly more complicated challenge that is only just starting to be attempted.

Innovative new approaches will emerge in the next few years for overcoming these challenges in enabling useful robotic behaviours in real environments. The next few years will see an interesting fusion of conventional approaches, such as hand-crafted robot control algorithms, with new approaches, like machine learning of robot behaviours.

Human-robot interaction, and human-AI collaborative working will also become increasingly important. What we are working towards now is “shared control” in which artificial intelligence and humans collaborate together to control a remote robot. For example, recent research from the French institute CNRS shows human operators controlling the motion of a robot arm and gripper in relation to an object, while the AI automatically orients the gripper with the object. Simultaneously, the AI automatically controls a second robot arm to move a camera so that it always offers the best viewpoint. Thus, human and AI collaborate to control a pair of arms for a remote handling task. In the future, this kind of human-machine collaboration could transform how we interact with nuclear and hazardous environments.

The Royal Society is currently carrying out a major policy project on machine learning, and its impact for the UK economy and society. To find out more please visit our project page.

This post was originally published by the Royal Society. Click here to view the original post.

The many potential social and economic benefits from advances in AI-based technologies depend entirely on the environment in which these technologies evolve, says the Royal Society. According to a new report from the UK’s science academy, urgent consideration needs to be given to the “careful stewardship” needed over the next ten years to ensure that the dividends from machine learning – the form of artificial intelligence that allows machines to learn from data – benefit all in UK society.

Machine Learning: the power and promise of computers that learn by example, published today (25 April 2017), comes at a critical time in the rapid development and use of this technology, and the growing debate about how it will reshape the UK economy and people’s lives. Crucially the report calls for research funding bodies to support a new wave of machine learning research that goes beyond technical challenges, and into areas aimed at addressing public confidence in machine learning – vital to the UK maintaining its internationally competitive edge at the forefront of this area.

The report also offers the first evidence about the UK public’s views on machine learning, including the application areas about which they are particularly positive, and the need for the real-world data feeding the growth of this technology to be dealt with fairly and securely.

Professor Peter Donnelly FRS, chair of the report’s working group and Director of the Wellcome Trust Centre for Human Genetics and Professor of Statistical Science at the University of Oxford, says:

“Machine learning is already used in many apps and services that we encounter every day. It is used to tag people in our photos, by our phones to interpret voice commands, by internet retailers to make recommendations, and by banks to spot unusual activity on a credit or debit card. However, these current applications only scratch the surface of understanding just how powerful a technology this could be.

“Machine learning will have an increasing impact on our lives and lifestyles over the next five to ten years. There is much work to be done so that we take advantage of machine learning’s potential and ensure that the benefits are shared, especially as this could be a key area of opportunity for the UK in the coming years.”

An action plan so no one is left behind and the benefits are shared

The report calls for action in a number of key areas over the next five to ten years to create an environment of “careful stewardship” that can help ensure that the benefits of this technology are felt broadly. Understanding who will be most affected, how the benefits are likely to be distributed, and where the opportunities for growth lie, will be key to designing the most effective interventions to enable people and businesses to adapt to, and take advantage of, the machine learning driven changes to their lives and livelihoods.

Supporting the development of skills at every level

• 23% of the UK population lack basic digital skills
• Build digital skills and understanding at every level from schools to universities, and into the workplace, and ensure that opportunities are not limited by gender, ethnicity or socio-economic background
• Government should consider introducing funded Masters courses in Machine Learning to develop a pool of informed users of machine learning across business, industry, and research sectors
• There is a critical need for increased training at PhD level and beyond to invest in the next generation of research leaders in machine learning.

Creating opportunities to use machine learning

• Integrate machine learning into the UK’s industrial strategy, to help businesses make the most of its potential
• Support a new wave of research in machine learning, including in areas that can address social or ethical concerns.

Creating a data environment that supports machine learning

• 90% of the world’s data has been created within the last five years
• Continue to build on the UK’s track record of open data and safe sharing of data
• Where there is value in accessing data that cannot be open – for example medical data or commercially sensitive industry data – ensuring there are frameworks and agreements in place which facilitate appropriate data sharing in these circumstances.

An enabling governance environment

• Support an informed public debate about what we want machine learning to do, and how the benefits are distributed
• Develop a new framework for data governance – one that can keep pace with the challenge of the governance of data and its uses in the 21st century. While there may be specific questions about the use of machine learning in particular contexts, these should be handled in a sector-specific way.

Results from the UK’s first in-depth assessment of public views on machine learning – carried out by the Royal Society and Ipsos MORI – demonstrate that while most people have not heard the term ‘machine learning’ (only 9% have), the vast majority have heard about or used at least one of its applications. The public do not have a single view on machine learning – attitudes, positive or negative, vary depending on the circumstances in which machine learning is being used. For example, computers that could help with medical diagnoses, or could help relieve pressure on professionals in core services such as health and social care, education, and policing were viewed more favourably than computers that could make investments in the stock market.

As with most new technologies, some of those interviewed expressed concerns. These included: depersonalisation, or machine learning systems replacing valued human experiences; the potential impact of machine learning on employment; the potential for machine learning systems to cause harm, for example accidents in autonomous vehicles; and machine learning systems restricting choice, such as when directing consumers to specific products and services.

Ongoing public confidence will be central to realising the benefits that machine learning promises, and continued engagement between machine learning researchers and practitioners and the public will be important as the field develops. The report also calls on researchers to consider the wider impact of their work and to receive training in recognising the ethical implications.

Most applications of deep learning use “convolutional” neural networks, in which the nodes of each layer are clustered, the clusters overlap, and each cluster feeds data to multiple nodes (orange and green) of the next layer. Image: Jose-Luis Olivares/MIT

In the past 10 years, the best-performing artificial-intelligence systems — such as the speech recognizers on smartphones or Google’s latest automatic translator — have resulted from a technique called “deep learning.”

Deep learning is, in fact, a new name for an approach to artificial intelligence called neural networks, which have been going in and out of fashion for more than 70 years. Neural networks were first proposed in 1944 by Warren McCullough and Walter Pitts, two University of Chicago researchers who moved to MIT in 1952 as founding members of what’s sometimes called the first cognitive science department.

Neural nets were a major area of research in both neuroscience and computer science until 1969, when, according to computer science lore, they were killed off by the MIT mathematicians Marvin Minsky and Seymour Papert, who a year later would become co-directors of the new MIT Artificial Intelligence Laboratory.

The technique then enjoyed a resurgence in the 1980s, fell into eclipse again in the first decade of the new century, and has returned like gangbusters in the second, fueled largely by the increased processing power of graphics chips.

“There’s this idea that ideas in science are a bit like epidemics of viruses,” says Tomas Poggio, the Eugene McDermott Professor of Brain and Cognitive Sciences at MIT, an investigator at MIT’s McGovern Institute for Brain Research, and director of MIT’s Center for Brains, Minds, and Machines. “There are apparently five or six basic strains of flu viruses, and apparently each one comes back with a period of around 25 years. People get infected, and they develop an immune response, and so they don’t get infected for the next 25 years. And then there is a new generation that is ready to be infected by the same strain of virus. In science, people fall in love with an idea, get excited about it, hammer it to death, and then get immunized — they get tired of it. So ideas should have the same kind of periodicity!”

Weighty matters

Neural nets are a means of doing machine learning, in which a computer learns to perform some task by analyzing training examples. Usually, the examples have been hand-labeled in advance. An object recognition system, for instance, might be fed thousands of labeled images of cars, houses, coffee cups, and so on, and it would find visual patterns in the images that consistently correlate with particular labels.

Modeled loosely on the human brain, a neural net consists of thousands or even millions of simple processing nodes that are densely interconnected. Most of today’s neural nets are organized into layers of nodes, and they’re “feed-forward,” meaning that data moves through them in only one direction. An individual node might be connected to several nodes in the layer beneath it, from which it receives data, and several nodes in the layer above it, to which it sends data.

To each of its incoming connections, a node will assign a number known as a “weight.” When the network is active, the node receives a different data item — a different number — over each of its connections and multiplies it by the associated weight. It then adds the resulting products together, yielding a single number. If that number is below a threshold value, the node passes no data to the next layer. If the number exceeds the threshold value, the node “fires,” which in today’s neural nets generally means sending the number — the sum of the weighted inputs — along all its outgoing connections.

When a neural net is being trained, all of its weights and thresholds are initially set to random values. Training data is fed to the bottom layer — the input layer — and it passes through the succeeding layers, getting multiplied and added together in complex ways, until it finally arrives, radically transformed, at the output layer. During training, the weights and thresholds are continually adjusted until training data with the same labels consistently yield similar outputs.

Minds and machines

The neural nets described by McCullough and Pitts in 1944 had thresholds and weights, but they weren’t arranged into layers, and the researchers didn’t specify any training mechanism. What McCullough and Pitts showed was that a neural net could, in principle, compute any function that a digital computer could. The result was more neuroscience than computer science: The point was to suggest that the human brain could be thought of as a computing device.

Neural nets continue to be a valuable tool for neuroscientific research. For instance, particular network layouts or rules for adjusting weights and thresholds have reproduced observed features of human neuroanatomy and cognition, an indication that they capture something about how the brain processes information.

The first trainable neural network, the Perceptron, was demonstrated by the Cornell University psychologist Frank Rosenblatt in 1957. The Perceptron’s design was much like that of the modern neural net, except that it had only one layer with adjustable weights and thresholds, sandwiched between input and output layers.

Perceptrons were an active area of research in both psychology and the fledgling discipline of computer science until 1959, when Minsky and Papert published a book titled “Perceptrons,” which demonstrated that executing certain fairly common computations on Perceptrons would be impractically time-consuming.

“Of course, all of these limitations kind of disappear if you take machinery that is a little more complicated — like, two layers,” Poggio says. But at the time, the book had a chilling effect on neural-net research.

“You have to put these things in historical context,” Poggio says. “They were arguing for programming — for languages like Lisp. Not many years before, people were still using analog computers. It was not clear at all at the time that programming was the way to go. I think they went a little bit overboard, but as usual, it’s not black and white. If you think of this as this competition between analog computing and digital computing, they fought for what at the time was the right thing.”

Periodicity

By the 1980s, however, researchers had developed algorithms for modifying neural nets’ weights and thresholds that were efficient enough for networks with more than one layer, removing many of the limitations identified by Minsky and Papert. The field enjoyed a renaissance.

But intellectually, there’s something unsatisfying about neural nets. Enough training may revise a network’s settings to the point that it can usefully classify data, but what do those settings mean? What image features is an object recognizer looking at, and how does it piece them together into the distinctive visual signatures of cars, houses, and coffee cups? Looking at the weights of individual connections won’t answer that question.

In recent years, computer scientists have begun to come up with ingenious methods for deducing the analytic strategies adopted by neural nets. But in the 1980s, the networks’ strategies were indecipherable. So around the turn of the century, neural networks were supplanted by support vector machines, an alternative approach to machine learning that’s based on some very clean and elegant mathematics.

The recent resurgence in neural networks — the deep-learning revolution — comes courtesy of the computer-game industry. The complex imagery and rapid pace of today’s video games require hardware that can keep up, and the result has been the graphics processing unit (GPU), which packs thousands of relatively simple processing cores on a single chip. It didn’t take long for researchers to realize that the architecture of a GPU is remarkably like that of a neural net.

Modern GPUs enabled the one-layer networks of the 1960s and the two- to three-layer networks of the 1980s to blossom into the 10-, 15-, even 50-layer networks of today. That’s what the “deep” in “deep learning” refers to — the depth of the network’s layers. And currently, deep learning is responsible for the best-performing systems in almost every area of artificial-intelligence research.

Under the hood

The networks’ opacity is still unsettling to theorists, but there’s headway on that front, too. In addition to directing the Center for Brains, Minds, and Machines (CBMM), Poggio leads the center’s research program in Theoretical Frameworks for Intelligence. Recently, Poggio and his CBMM colleagues have released a three-part theoretical study of neural networks.

The first part, which was published last month in the International Journal of Automation and Computing, addresses the range of computations that deep-learning networks can execute and when deep networks offer advantages over shallower ones. Parts two and three, which have been released as CBMM technical reports, address the problems of global optimization, or guaranteeing that a network has found the settings that best accord with its training data, and overfitting, or cases in which the network becomes so attuned to the specifics of its training data that it fails to generalize to other instances of the same categories.

There are still plenty of theoretical questions to be answered, but CBMM researchers’ work could help ensure that neural networks finally break the generational cycle that has brought them in and out of favor for seven decades.

YouTube: Dylan F Glas

“Robot’s Delight – A Lyrical Exposition on Learning by Imitation from Human-Human Interaction” is a video submission that won Best Video at the 2017 ACM/IEEE International Conference on Human-Robot Interaction (HRI 2017). The team also provides an in-depth explanation of the techniques and robotics in the video.

Watch it here:

Select video clips (c) 2016, IEEE. Reused, with permission.

Although social robots are growing in popularity and technical feasibility, it is still unclear how we can effectively program social behaviors. There are many difficulties in programming social robots — we need to design hundreds or thousands of dialogue rules, anticipate situations the robot will face, handle common recognition errors, and program the robot to respond to many variations of human speech and behavior. Perhaps most challenging is that we often do not understand the reasoning behind our own behavior and so it is hard to program such implicit knowledge into robots.

In this video, we present two studies exploring learning-by-imitation from human-human interaction. In these studies, we developed techniques for learning typical actions and execution logic directly from people interacting naturally with each other. We propose that this is more scalable and robust than developing interaction logic by hand, and it requires much less effort.

In the first study, we asked participants to role-play a shopkeeper and a customer in a camera shop scenario, and we recorded their motion and speech in 178 interactions. By extracting typical motion and speech actions using unsupervised clustering algorithms, we created a set of robot behaviors and trained a machine learning classifier to predict which of those actions a human shopkeeper would have performed in any given situation. Using this classifier, we programmed a Robovie robot to imitate the movement and speech behavior of the shopkeeper, e.g. greeting the customer, answering questions about camera features, and introducing different cameras.

Experiment results showed that our techniques enabled the robot to perform correct behaviors 84.8% of the time, which was particularly interesting since speech recognition was only 76.8% accurate. This illustrates the robustness of the system to sensor noise, which is one advantage of using noisy, real-world data for training.

In the second study, we used a similar technique to train the android ERICA to imitate people’s behavior in a travel agent scenario. In this case, the challenge was to model the topic of the interaction so that ERICA could learn to answer ambiguous questions like “how much does this package cost?”.

We did this by observing that utterances in the same topic tend to occur together in interactions, so we calculated co-occurrence metrics similar to those used in product recommendation systems for online shopping sites. Using these metrics, we were able to cluster the customer and shopkeeper actions into topics, and these topics were used to improve ERICA’s predictor in order to answer ambiguous questions.

In both of these studies, we illustrated a completely hands-off approach to developing robot interaction logic – the robots learned only from example data of people interacting with each other, and no designer or programmer was needed! We think scalable, data-driven techniques like these promise to be powerful tools for developing even richer, more humanlike interaction logic for robots in the future.

The extended abstract for this video can be found here.

For the full details of the Robovie study, please see our IEEE Transactions on Robotics paper.

Phoebe Liu, Dylan F. Glas, Takayuki Kanda, and Hiroshi Ishiguro, Data-Driven HRI: Learning Social Behaviors by Example from Human-Human Interaction, in IEEE Transactions on Robotics, Vol. 32, No. 4, pp. 988-1008, 2016.

Author’s preprint is here.

Artificial intelligence (AI) already plays a major role in human economies and societies, and it will play an even bigger role in the coming years. To ponder the future of AI is thus to acknowledge that the future is AI.

This will be partly owing to advances in “deep learning,” which uses multi-layer neural networks that were first theorized in the 1980s. With today’s greater computing power and storage, deep learning is now a practical possibility, and a deep-learning application gained worldwide attention in 2016 by beating the world champion in Go. Commercial enterprises and governments alike hope to adapt the technology to find useful patterns in “Big Data” of all kinds.

In 2011, IBM’s Watson marked another AI watershed, by beating two previous champions in Jeopardy!, a game that combines general knowledge with lateral thinking. And yet another significant development is the emerging “Internet of Things,” which will continue to grow as more gadgets, home appliances, wearable devices, and publicly-sited sensors become connected and begin to broadcast messages around the clock. Big Brother won’t be watching you, but a trillion little brothers might be.

Image: CB Insights

From Jeopardy winners and Go masters to infamous advertising-related racial profiling, it would seem we have entered an era in which artificial intelligence developments are rapidly accelerating. But a fully sentient being whose electronic “brain” can fully engage in complex cognitive tasks using fair moral judgement remains, for now, beyond our capabilities.

Unfortunately, current developments are generating a general fear of what artificial intelligence could become in the future. Its representation in recent pop culture shows how cautious – and pessimistic – we are about the technology. The problem with fear is that it can be crippling and, at times, promote ignorance.

Learning the inner workings of artificial intelligence is an antidote to these worries. And this knowledge can facilitate both responsible and carefree engagement.

The core foundation of artificial intelligence is rooted in machine learning, which is an elegant and widely accessible tool. But to understand what machine learning means, we first need to examine how the pros of its potential absolutely outweigh its cons.

Data is the key

Simply put, machine learning refers to teaching computers how to analyse data for solving particular tasks through algorithms. For handwriting recognition, for example, classification algorithms are used to differentiate letters based on someone’s handwriting. Housing data sets, on the other hand, use regression algorithms to estimate in a quantifiable way the selling price of a given property.Machine learning, then, comes down to data. Almost every enterprise generates data in one way or another: think market research, social media, school surveys, automated systems. Machine learning applications try to find hidden patterns and correlations in the chaos of large data sets to develop models that can predict behaviour.

Credit: Jonathan Khoo/Flickr, CC

Machine learning, then, comes down to data. Almost every enterprise generates data in one way or another: think market research, social media, school surveys, automated systems. Machine learning applications try to find hidden patterns and correlations in the chaos of large data sets to develop models that can predict behaviour.

Data have two key elements – samples and features. The former represents individual elements in a group; the latter amounts to characteristics shared by them.

Look at social media as an example: users are samples and their usage can be translated as features. Facebook, for instance, employs different aspects of “liking” activity, which change from user to user, as important features for user-targeted advertising.

Facebook friends can also be used as samples, while their connections to other people act as features, establishing a network where information propagation can be studied.

My Facebook friends network: each node is a friend who might or might not be connected to other friends. The larger the node, the more connections one has. Similar colours indicate similar social circles. https://lostcircles.com/

Outside of social media, automated systems used in industrial processes as monitoring tools use time snapshots of the entire process as samples, and sensor measurements at a particular time as features. This allows the system to detect anomalies in the process in real time.

All these different solutions rely on feeding data to machines and teaching them to reach their own predictions once they have strategically assessed the given information. And this is machine learning.

Human intelligence as a starting point

Any data can be translated into these simple concepts and any machine-learning application, including artificial intelligence, uses these concepts as its building blocks.

Once data are understood, it’s time to decide what do to with this information. One of the most common and intuitive applications of machine learning is classification. The system learns how to put data into different groups based on a reference data set.

This is directly associated with the kinds of decisions we make every day, whether it’s grouping similar products (kitchen goods against beauty products, for instance), or choosing good films to watch based on previous experiences. While these two examples might seem completely disconnected, they rely on an essential assumption of classification: predictions defined as well-established categories.

When picking up a bottle of moisturiser, for example, we use a particular list of features (the shape of the container, for instance, or the smell of the product) to predict – accurately – that it’s a beauty product. A similar strategy is used for picking films by assessing a list of features (the director, for instance, or the actor) to predict whether a film is in one of two categories: good or bad.

By grasping the different relationships between features associated with a group of samples, we can predict whether a film may be worth watching or, better yet, we can create a program to do this for us.

But to be able to manipulate this information, we need to be a data science expert, a master of maths and statistics, with enough programming skills to make Alan Turing and Margaret Hamilton proud, right? Not quite.

You don’t have to be Alan Turing to have a go at machine learning. Credit: CyberHades/Flickr, CC

We all know enough of our native language to get by in our daily lives, even if only a few of us can venture into linguistics and literature. Maths is similar; it’s around us all the time, so calculating change from buying something or measuring ingredients to follow a recipe is not a burden. In the same way, machine-learning mastery is not a requirement for its conscious and effective use.

Yes, there are extremely well-qualified and expert data scientists out there but, with little effort, anyone can learn its basics and improve the way they see and take advantage of information.

Algorithm your way through it

Going back to our classification algorithm, let’s think of one that mimics the way we make decisions. We are social beings, so how about social interactions? First impressions are important and we all have an internal model that evaluates in the first few minutes of meeting someone whether we like them or not.

Two outcomes are possible: a good or a bad impression. For every person, different characteristics (features) are taken into account (even if unconsciously) based on several encounters in the past (samples). These could be anything from tone of voice to extroversion and overall attitude to politeness.

For every new person we encounter, a model in our heads registers these inputs and establishes a prediction. We can break this modelling down to a set of inputs, weighted by their relevance to the final outcome.

For some people, attractiveness might be very important, whereas for others a good sense of humour or being a dog person says way more. Each person will develop her own model, which depends entirely on her experiences, or her data.

Different data result in different models being trained, with different outcomes. Our brain develops mechanisms that, while not entirely clear to us, establish how these factors will weigh out.

What machine learning does is develop rigorous, mathematical ways for machines to calculate those outcomes, particularly in cases where we cannot easily handle the volume of data. Now more than ever, data are vast and everlasting. Having access to a tool that actively uses this data for practical problem solving, such as artificial intelligence, means everyone should and can explore and exploit this. We should do this not only so we can create useful applications, but also to put machine learning and artificial intelligence in a brighter and not so worrisome perspective.

There are several resources out there for machine learning although they do require some programming ability. Many popular languages tailored for machine learning are available, from basic tutorials to full courses. It takes nothing more than an afternoon to be able to start venturing into it with palpable results.

All this is not to say that the concept of machines with human-like minds should not concern us. But knowing more about how these minds might work will gives us the power to be agents of positive change in a way that can allow us to maintain control over artificial intelligence and not the other way around.

This article was originally published on The Conversation. Read the original article.

If you liked this article, you may also want to read:

• Building character AI through machine learning
• What is machine learning?
• Brian Cox presents Science Matters: Machine learning and artificial intelligence
• Talking Machines: Machine learning and the Flint water crisis, with Jake Abernethy
• Talking Machines: Computational learning theory and machine learning for understanding cells, with Aviv Regev

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

Science fiction has already explored the theme of robot rights, such as the film Bicentennial Man. Image: Columbia Pictures

Science fiction likes to depict robots as autonomous machines, capable of making their own decisions and often expressing their own personalities. Yet we also tend to think of robots as property, and as lacking the kind of rights that we reserve for people.

But if a machine can think, decide and act on its own volition, if it can be harmed or held responsible for its actions, should we stop treating it like property and start treating it more like a person with rights?

What if a robot achieves true self-awareness? Should it have equal rights with us and the same protection under the law, or at least something similar?

These are some of the issues being discussed by the European Parliament’s Committee on Legal Affairs. Last year it released a draft report and motion calling for a set of civil law rules on robotics regulating their manufacture, use, autonomy and impact upon society.

Of the legal solutions proposed, perhaps most interesting was the suggestion of creating a legal status of “electronic persons” for the most sophisticated robots.

Approaching personhood

The report acknowledged that improvements in the autonomous and cognitive abilities of robots makes them more than simple tools, and makes ordinary rules on liability, such as contractual and tort liability, insufficient for handling them.

For example, the current EU directive on liability for harm by robots only covers foreseeable damage caused by manufacturing defects. In these cases, the manufacturer is responsible. However, when robots are able to learn and adapt to their environment in unpredictable ways, it’s harder for a manufacturer to foresee problems that could cause harm.

The report also questions about whether or not sufficiently sophisticated robots should be regarded as natural persons, legal persons (like corporations), animals or objects. Rather than lumping them into an existing category, it proposes that a new category of “electronic person” is more appropriate.

The report does not advocate immediate legislative action, though. Instead it proposes that legislation be updated if robots become more complex; if and when they develop more behavioural sophistication. If this occurs, one recommendation is to reduce the liability of “creators” proportional to the autonomy of the robot, and that a compulsory “no-fault” liability insurance could cover the shortfall.

But why go so far as to create a new category of “electronic persons”? After all, computers still have a long way to go before they match human intelligence, if they ever do.

But it can be agreed that robots – or more precisely the software that controls them – is becoming increasingly complex. Autonomous (or “emergent”) machines are becoming more common. There are ongoing discussions about the legal liability for autonomous vehicles, or whether we might be able to sue robotic surgeons.

These are not complicated problems as long as liability rests with the manufacturers. But what if manufacturers cannot be easily identified, such as if open source software is used by autonomous vehicles? Whom do you sue when there are millions of “creators” all over the world?

Artificial intelligence is also starting to live up to its moniker. Alan Turing, the father of modern computing, proposed a test in which a computer is considered “intelligent” if it fools humans into believing that the computer is human by its responses to questions. Already there are machines that are getting close to passing this test.

There are also other incredible successes, such as the computer that creates soundtracks to videos that are indistinguishable from natural sounds, the robot that can beat CAPTCHA, one that can create handwriting indistinguishable from human handwriting and the AI that recently beat some of the world’s best poker players.

Robots may eventually match human cognitive abilities and they are becoming increasingly human-like, including the ability to “feel” pain.

If this progress continues, it may not be long before self-aware robots are not just a product of fantastic speculation.

The EU report is among the first to formally consider these issues, but other countries are also engaging. Peking University’s Yueh-Hsuan Weng writes that Japan and South Korea expect us to live in a human-robot coexistence by 2030. Japan’s Ministry of Economy, Trade and Industry has created a series of robot guidelines addressing business and safety issues for next generation robots.

Electronic persons

If we did give robots some kind of legal status, what would it be? If they behaved like humans we could treat them like legal subjects rather than legal objects, or at least something in between. Legal subjects have rights and duties, and this gives them legal “personhood”. They do not have to be physical persons; a corporation is not a physical person but is recognised as a legal subject. Legal objects, on the other hand, do not have rights or duties although they may have economic value.

Assigning rights and duties to an inanimate object or software program independent of their creators may seem strange. However, with corporations we already see extensive rights and obligations given to fictitious legal entities.

Perhaps the approach to robots could be similar to that of corporations? The robot (or software program), if sufficiently sophisticated or if satisfying certain requirements, could be given similar rights to a corporation. This would allow it to earn money, pay taxes, own assets and sue or be sued independently of its creators. Its creators could, like directors of corporations, have rights or duties to the robot and to others with whom the robot interacts.

Robots would still have to be partly treated as legal objects since, unlike corporations, they may have physical bodies. The “electronic person” could thus be a combination of both a legal subject and a legal object.

The European Parliament will vote on the resolution this month. Regardless of the result, reconsidering robots and the law is inevitable and will require complex legal, computer science and insurance research.

This article was originally published on The Conversation. Read the original article.

An Air Force airman at the control module of an MQ-9 Reaper. Credit: John Bainter/USAF

Machine ethics offers an alternative solution for artificial intelligence (AI) safety governance. In order to mitigate risks in human-robot interactions, robots will have to comply with humanity’s ethical and legal norms, once they’ve merged into our daily life with highly autonomous capability. In terms of technical challenges, there are still many open questions in machine ethics. For example, what is deontic logic and how can it be used for improving AI safety? How do we fashion the knowledge representation for ethical robots? What about the adaptiveness of the ethical governance? These are all significant questions for us to investigate.

In this interview, we invite Prof. Ronald C. Arkin to share his insights on robot ethics, with a focus on its technical aspects.

Date: February 1st 2017

Interviewee: Prof. Ronald C. Arkin, Regents’ Professor and Associate Dean in the College of Computing at the Georgia Institute of Technology and Executive Committee Member of IEEE Global Initiative for Ethical Considerations in Artificial Intelligence and Autonomous Systems

Interviewer: Dr. Yueh-Hsuan Weng, Co-founder of ROBOLAW.ASIA and Fellow at CLAST AI & Law Research Committee (Beijing) and Tech and Law Center (Milan).

Ron Arkin

Q1 WENG: Thank you for agreeing to an interview with us. Can you tell us a little about your background?

ARKIN: I have been a roboticist for 35 years, so it’s hard to tell [only] a little, but I will do my best. Currently I am a Regents’ Professor and Associate Dean in the College of Computing at the Georgia Institute of Technology. I got my Ph.D. from the University of Massachusetts in 1987, and have been working on robotics since that time. I actually have degrees in chemistry as well—a bachelor’s and a master’s in organic chemistry. So, I have a rather diverse background and followed a non-traditional path compared to most computer scientists.

Q2 WENG: What is your definition of “Ethical Robots”?

ARKIN: That’s an interesting question! I have been asked that one before. I would say an ethical robot is a robotic system that is able to comply with the ethical and social norms of human beings—in some cases in accordance with international law, in other cases in accordance with national or local law. In addition, also being able to comport with humans in a natural manner, obey, and adhere to the same principles as an ethical person.

Q3 WENG: Could you please tell us why you are particularly interested in “Embedded Ethics”? Why is this concept so important?

ARKIN: That’s why you need to read the preface of my book [Governing Lethal Behaviour in Autonomous Robots]. It’s a rather long story. But let me try to give you a short version. Again, it dates back to a meeting—the first international symposium in Roboethics in 2004—where I was exposed to international conventions, e.g., the Geneva convention. It also led me to understand the Vatican’s position on how they viewed appropriate ethical interactions. Perhaps even more important than that one experience was the fact that we roboticists that have been in the field for some time finally started to have success, and we were seeing our technology move out of the laboratory community and into the real world. As such, it wasn’t hard to see how this could go in a variety of different directions.

Some robotics technology could benefit humanity, some not so much. At that point, one has to ask, “who is responsible for this?” So I assumed that, in part, it was myself. I believe that I have to take the responsibility for the technology that I helped to create. The issue is that we, as roboticists, had actually succeeded in many ways in terms of creating technology that we only envisioned when we first started that is going to have potentially profound impact on our society and the world as a whole. As such, we should be proactive in the ways in which we move forward [in order] to get a better understanding of the consequences of the work, and the social implications of the technology we are designing.

https://www.youtube.com/watch?v=jOJwkvihYtc

Q4 WENG: The IEEE Global Initiative for Ethical Considerations in Artificial Intelligence and Autonomous Systems has published their report “Ethically Aligned Design” in December 2016. As a member of the IEEE Global Initiative, could you please tell us about your role?

ARKIN: I serve a variety of roles. First, I am on the executive committee of that particular initiative, so I work at a high-level with about 30 other people or so. We have committees that large. I also co-chair the committee in Affective Computing. Joanna Bryson is my co-chair on that.

Affective Computing is concerned with the issues surrounding the illusion of emotions we can potentially create in these artifacts, and the ways human beings can bond to them. We are exploring a variety of different things, including robot intimacy, or the notion of applying robotic nudges and moving people in certain directions subtly, and in so doing to change the way people act and behave using these technologies, while trying to understand what may or may not be appropriate guidelines based on ethical considerations.

It’s about 15-20 people in that committee. We are one of the newer committees and we only have a couple of pages in the initial design report. But actually, in the June meeting in Austin, for the next release of the document 2.0, we will have as much as the other committees do. Finally, I am also a member of the humanitarian committee. I think it is called economic and humanitarian. I don’t know the exact title, but I am interested in the humanitarian side, and potential uses of robotics for humanitarian applications.

Q5 WENG: What is “Deontic Logic”? Why does it matter to machine ethics?

ARKIN: It may or may not matter. It’s just one approach to a particular problem. Deontic logic is a logic of prohibitions, obligations and permissions. It is one type of framework that is more transparent in order to get a better understanding of the way in which systems operate in the real world. Then it explains its reasoning to a human being, as compared to say deep learning or other forms of neural networks which may indeed provide good results but provide no explanatory capability to humans after the decision has been made.

The notion of prohibitions basically tells the robot some things it cannot do. The notion of permissions describe the things it is potentially allowed to do. And the notion of obligations says things that the robot must do under specific circumstances. It provides a nice framework of being able to explore morality. Others have already explored this, like Bringsjord at RPI who has applied deontic logic in this framework. We are certainly not the first in that particular space. Much of the work we have done has dealt with first-order predicate logic as a basis. It doesn’t have to be deep because it is so explicit in many cases.

Q6 WENG: Your hybrid embedded ethics framework aims for governing lethal behavior of military robots. Can we apply the same technical framework to service robots? For example, personal care robots might have some non-lethal, but still risky autonomous behaviors that need to be regulated by machine ethics.

ARKIN: I believe so. Indeed the whole point of the architecture is trying to design the systems to make them generalizable and not ad hoc for particular sets of circumstances. We have yet to look at service robotics in a traditional sense, but we have already developed something we refer to as the intervening Ethical Governor, which we are using for health care applications—particularly for assisting the management of two human beings in a particular relationship, like a patient and their care giver.

A robot hands a medication bottle to a person. Photo credit: Keith Bujak. Source: Georgia Tech News Center

We are doing work for the National Science Foundation, joint with Tufts University, looking at how in early stage Parkinson’s there is a condition called the facial masking. The loss of muscle movements in the face of the patient can produce stigmatization, and the patient-care giver relationship deteriorates due to the loss of non-verbal communication. We just had a paper accepted at ICRA describing the architecture that we have and how it has been modified from the military version. We generalized it from the military constraint-based version to a rule-based version.

But also, instead of using the international humanitarian law, we are using occupational therapy guides as a basis for generating the responses when two human beings in the relationship step outside of acceptable bounds. So the acceptable bounds now are not defined by international humanitarian law, but they are defined by occupational therapy—such as when a patient or a care giver start to yell at each other, or one starts to leave the room. The hard part is recognizing when dignity is being abridged, particularly in the patient. So we model the moral emotions of empathy, shame and embarrassment. We tried to ensure that the empathy level in the care giver remains within acceptable bounds and that the patient shame and embarrassment doesn’t rise to unacceptable levels. So the intervening Ethical Governor intervenes when a really inappropriate event occurs. It then provides strong corrective action by the robot to encourage people to get along better. The moral adaptor component which we used in the ethical architecture—initially with the guilt model—has now been extended to other moral emotions like shame, embarrassment, guilt and empathy [in order to] try and provide more subtle non-verbal queues [to help the] robot steer the relationship in the correct way before it get out of hand.

Q7 WENG: Regarding your research on machine ethics, what kind of indicators you use to evaluate the final outputs of your project?

ARKIN: In the first one, we were designing a prototype to prove the concept of the architecture. We never make the claim that this is the best way to do it. It is just one way to do it. Then we demonstrate that particular capability through the use of simulations and other aspects. The system was not intended to be fielded, but hopefully others could take those core ideas and potentially put them to use in the real world if they so choose. For the intervening governor, we’re running human subject tests right now—not with Parkinson’s patients yet, but with the elderly population, 50 or older. We’re trying to get a better understanding of the correct sets of rules, and to correct the interventions in a normative fashion, so we can ultimately move on in the final year of the grant to work with actual Parkinson’s patients. Again, using standard HRI technics, such as recording the kinds of reactions they have and parsing the data from the sessions and using pre- and post- evaluation questionnaires.

Q8 WENG: Let’s talk about your proposed machine ethics framework. You had mentioned that it includes four main parts: (1) Ethical Governor (2) Ethical Behavior Control (3) Ethical Adaptor (4) Responsibility Advisor. First of all, what is the major difference between Ethical Governor and Ethical Behavior Control? Is it the difference between deliberative and reactive control, or something else?

ARKIN: The Ethical Governor is basically a standalone component to the architecture which is a bolt on component at the end of the existing system. In the case of military scenario: asking for permission—should it fire or not. This is an intervening stage which either grants or denies permission through the validation of permission-to-fire variable. So it is straightforward thing to be able to do. The Behavior Ethical Control is a different kind of thing as it works purely with the reactive component. What it tries to do is, for each of the individual behaviors themselves, to respond whether or not to engage a target. This is prior to dealing with the Ethical Governor making the decision; trying to make the decision correctly before it gets to the Ethical Governor, where individual behaviors are given ethical guidelines to try and ensure that the behavior is correct itself. Each of the behaviors is provided with appropriate ethical constraints. So the difference is: one is located within the reactive component in each of the individual behaviors that makes the activity of the architecture, and the Ethical Governor is at the end of the architectural chain.

Q9 WENG: What is the importance of Ethical Adaptor? What kinds of machine learning skill that you chose to improve its adaptiveness?

ARKIN: Machine learning is something we haven’t generally explored in this particular context. Especially the military context, we kind of engineer out the adaptiveness and machine learning for fear that the robot would be granted the authority to start to generalize the targets in the field, which we feel is inappropriate. Learning after action is okay. but that’s a different story. The importance of the Ethical Adaptor, and its cognitive models of guilt that we use, is the fact that even the actions the agent takes may inadvertently end up unethical. This is referred to as a “moral luck”. It could be good luck or bad luck. But if it is bad luck, you want to be able to ensure proactively that the bad action doesn’t occur again. That’s actually the functions of cognitive guilt for people as well to reduce the likelihood of reoccurrence of a behavior that resulted in bad actions. By giving this to a robot we hope that if it makes a mistake it will not repeat it. The probability of that behavior reoccurring would be reduced.

The best example is the case of battle damage assessment for a military scenario. For example, if a robot does its assessment and it looks like a released weapon wouldn’t create too much collateral damage, within acceptable limits, then it will release the weapon. But after it does it, the battle damage action is actually more severe then, it initially appeared to be. Under those circumstances, we want the robot in the field to reduce the probability of releasing the weapon again under similar circumstances. That’s, in essence, the analogue of guilt we use, and eventually, it will get to the point that the robot will not engage targets at all. Again, the software is updated to make sure the use is consistent with Geneva convention.

Q10 WENG: The Laws of War (LOW) and Rules of Engagement (ROE) are written in the form of natural language. How did you transfer these norms into machine readable formats?

ARKIN: It’s important to keep it in mind that we do not try and capture all the Rules of Engagement and all the Laws of War. We always advocate that the system should be used in very narrow circumstances and very narrow situations, such as building clearing operations or counter-sniper operations. It’s impossible to capture all human morality and all the Laws of War at this particular point of time. So, the question is what are the relevant Laws of War and Rules of Engagement for these particular narrow situations that the robot will find itself in.

Let’s take a room clearing operation. We try and do this using these constraint-based approaches. The Laws of War basically say that if you see a civilian, you should not shoot them; if you see a surrendered person, you should not shoot them. For aerial vehicles, if you are outside the kill zone (this is the simulations we did) which is designated by the Rules of Engagement, you cannot fire your weapon—explain simply why you cannot do that. If you release the weapon a toward a legitimate target but it would result in damage of cultural property, like a mosque or church or some similar circumstance, that is illegal as well. You are forbidden from doing those sort of things.

So, the initial work we did was more for aerial circumstances. We also computed the collateral damage using a battle damage assessment strategy to determine whether if it is within acceptable limits for a particular target and to restrain the weapon for proportionality. To be able to do this you have to be able to understand the military necessity of the target, which is assigned by a human being. You have to be able to understand the proportionality of the weapon. In other words: Which weapon is sufficient to take out the target but to do no more damage of the surroundings than is absolutely necessary? You have to ensure that you don’t destroy cultural property as well. Basically, adhere to all necessary principles, such as discrimination which is to be able to tell apart combatants and non-combatants. We also state we should restrict the use of these particular systems for high-intensity combat operations, or inter-state warfare, and not the kind of circumstances you find in modern counter- insurgency operations. They should be used with great care and great caution in very narrow circumstances where the predictability of their performance is far more likely to be good.

Q11 WENG: Why did you choose case-based reasoning (CBR) but not rule-based reasoning (RBR) when you designed Responsibility Advisor?
ARKIN: Historically we used case-based reasoning a lot for a variety of different reasons. Part of the reason is to gain information from previous experts. One of the examples in the past was to work with snipers in the laboratory. We can work with experts. It can be legal experts, it can be medical experts, it can be military experts. Again, the primary reason is to be able to readily use the experience of warfighters to guide us in operations.

Q12 WENG: Finally, your proposal “Governing Lethal Behavior: Embedding Ethics in a Hybrid Deliberative/Reactive Robot Architecture” was published in 2007. There have been great technological improvements in artificial intelligence in the past decade. Are there any recent updates from your robot ethics research, or anything that you want to share with us?

ARKIN: The project ended around 2010, so there is additional information. As I say, I would encourage those who are really interested to take a look at the book entitled “Governing Lethal Behavior in Autonomous Robots” . That represents the final statement and outcome for that particular project and includes the significant technical details describing that work. Recent advances include the application in other domains, such as the caregiver-patient relationship [discussed above] using the same underlying principles, including both the Moral Adaptor and Ethical Governor in a slightly different. If we were successful, it may also be useful for assisting other human-human relationships, such as a parent and a teenager who may be at odds with one another or even in marriage consulting. We hope to use robots for the benefit of mankind, whether it’s to make sure that warfare is conducted within internationally agreed upon laws, or whether it’s to ensure that people behave appropriately toward each other

If you enjoyed this article, you may also enjoy:

• How do we regulate robo-morality?
• One being for two origins: A new perspective on roboethics
• Considering robot care, ethics & future tech at #ERW2016 European Robotics Week
• On the ethics of research in robotics, with Raja Chatila
• A European perspective on robot law: Interview with Mady Delvaux-Stehres
• Open Roboethics initiative delivers statement to United Nations CCW

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

bots_alive Gotcha

Modern character AI could be better.

I’m unaware of any NPCs or electronic toy characters that can sustain an illusion of life over more than an hour. They suffer from predictability, simplicity, and inauthentic verbal and nonverbal behavior.

This post is about a new form of character AI. Much like motion capture for scripted animation, this new technique may revolutionize how interactive characters are created, through observation of authentic human-generated behavior.

I’ll explain via a product on which we’re pioneering this technique.

Bots_alive! A robot that plays and finds its way

We launched bots_alive last week on Kickstarter (campaign here). Here’s a 1-minute teaser video with mostly scripted motion.

The product is a smartphone kit that gives simple but lifelike autonomy to a popular RC toy, the Hexbug® Spider.

That video shows the product as we envision it at shipping time. Here you can see how it acts now, with this new character AI technique driving its core functionality already.

Training data to build a character

The character AI for these robots is created in a new way. Here’s how it works.

A human teleoperates the robot through many different scenarios. The teleoperator acts as an improvising puppeteer; he or she follows heuristics yet also empathizes with the robot and authentically responds to what happens, whether it be with curiosity, fear, jubilation, or another emotion.

For bots_alive, the teleoperator watches a screen display of what the system sees and pushes buttons to send commands of forward, back, left, right, forward-right, forward-left, back-right, or back-left. We also consider not pushing a button to be a no-action command.

From these teleoperation sessions, we gather training data containing (a) teleoperation commands and (b) information about the context in which each command was given. A category of machine learning called supervised learning is applied to create a model of the puppeteer, which effectively answers the question: In context X, what is the probability that the teleoperator would have given command Y?

Creating a behavioral model

All of that happens during development. In the end user’s hands, the robot is autonomously controlled by the model. The model of the puppeteer is the character AI.

Let’s assume that the contextual information in the training data is sufficient to justify why the teleoperator chooses one action over another. Then the data contains the spontaneity, uncertainty, and social authenticity* expressed by the human puppeteer. And the autonomous robot retains these qualities if the machine learning is effective.

Why go to the trouble of using machine learning?

We expect this process to create character AI that feels more organic and alive than traditional character AI.

Traditional methods consist of imagining and writing rules for behavior or a finite state machine. It’s much more of an abstract exercise than immersing someone in a situation as the puppeteer.

If you play video games, imagine how much you sometimes empathize with the character you’re controlling. You may even forget the separation between the two of you, experiencing the world as that character. Consider whether your control in these moments is different, both at a high level and in tiny movements, than what the character would do if you had simply written down a set of rules for it to act by.

In that difference lies the promise of this method for developing character AI.

In psychology research methodology, there’s a broad consensus that if you want to know what someone would do in a situation, you don’t ask them what they would do. You put them in the situation and observe. Traditional character AI development is like asking the human developer or designer. Our method is instead analogous to observing a human immersed in the experience of the character. That’s why we expect it to create more authentic characters.

Our method is an application of what’s called learning from demonstration.** For times the teleoperation needs to include human interaction with the character, we keep the teleoperation secret to keep the human interaction partner from changing his or her behavior from what would be done with an autonomous character. We coined this secret version learning from the wizard (or LfW) when it was a research project at the MIT Media Lab. The name combines learning from demonstration with the Wizard of Oz experimental paradigm.

Training as interactive machine learning

In our development, I act as the teleoperator and machine learning programmer. The training process is not simply a bunch of demonstration. Rather, it’s an iterative process of:

• demonstrate
• apply machine learning on the demonstration data set
• observe behavior from the learned model
• create more demonstrations in those contexts the robot isn’t acting satisfactorily
• apply machine learning, and so on.

A general representation of interactive machine learning. From Amershi et al., 2015. Power to the People: The Role of Humans in Interactive Machine Learning. In AI Magazine.

Throughout these iterations, the teleoperator and algorithm designer also reflect on what contextual information still needs to be encoded to improve learning, determine what context cannot be encoded and therefore should be ignored by the teleoperator, and find fun and more delightful behavior than originally puppeteered.

How can we know that it’s better?

We can’t. Not with 100% confidence. But we have already seen that it can make compelling character AI.

We saw it in our study at MIT. In a relatively large randomized experiment, kids interacted either with a human-teleoperated robot or one with autonomous behavior learned from previous teleoperation. Compared to a no-robot condition, the teleoperated robot and the autonomous robot programmed by machine learning elicited similar behavior from their human interaction partners. Curiously, when children were asked whether the robot was human-controlled or autonomous, approximately half in each condition thought it was human-controlled. If you’re familiar with the Turing Test, the popularly known test of the effectiveness of an artificial intelligence, you might recognize that the MIT study constitutes passing a certain narrow and social Turing Test.

Compelling character AI was achieved in another such Turing test, using a technique similar to learning from demonstration, in Unreal Tournament 2004. BotPrize is a yearly academic competition for the bot that best passes as human. The 2012 winner was a bot by Jacob Schrum, Igor Karpov, and Risto Miikkulainen. It combined algorithmic optimization—which can result in markedly unnatural behavior that still achieves the specified goal— of its combat effectiveness with selective replay of recorded human behavior. Twenty-five human judges played alongside the Unreal-playing bots, repeatedly judging whether the other characters were human or artificial during combat. The bot by Schrum and colleagues was judged human more than 51.9% of the time. That percentage was just below the two human characters who were judged most human, at 53.3% and 52.2%. The algorithm is described here.

Lastly, we see compelling character AI now, in our playtesting of bots_alive.

The limitations of this approach are (1) sensing and encoding contextual information and (2) applying machine learning effectively. Both academia and industry are intensely researching how to improve those two limitations. Major strides have been made in the past 10 years, most notably through deep learning, and so it seems likely that further improvement is yet to come.

Our ability to write down behavioral rules that encode authentic behavior with an illusion of life? If there’s a reason why that will improve meaningfully, I don’t see it.

The webcast will take place on today from 9am-5:30pm EST and February 1st from 9am-5pm EST. Webcast participants are encouraged to submit questions for the presenters by e-mailing Michelle Schwalbe at mschwalbe@nas.edu who will read them out if time permits.

Machine learning is at the core of many applications that have become part of daily life, from voice recognition to image perception. These technologies, which a few years ago were performing at noticeably below-human levels, can now outperform people at some tasks. As the field continues to evolve, machine learning has the potential to play a transformative role across a diverse range of sectors including transportation, medicine, public services, and finance. This forum will bring together scientists from the UK and the US to explore potential applications for machine learning and discuss the legal and ethical challenges that could arise as machine learning algorithms are implemented.

Watch the live webcast here.

Session outlines below:

Session 1: The Frontiers of Machine Learning

The ubiquity of data, accessibility of computing power, and algorithmic advances have driven rapid progress in machine learning over the past five years. Not only does machine learning now underpin many applications that have become part of daily life, the field continues to evolve quickly, and has the potential to play a transformative role across a diverse range of sectors. This session will explore the frontiers of machine learning, in terms of both cutting-edge technology and near-term applications, and discuss the state of the art of machine learning.

Session 2: Machine learning and society

People and machine learning systems are increasingly interacting through a range of applications or contexts. This expansion of machine learning raises legal and ethical questions, re-frames discussions about uses of data, and poses new challenges for the governance of this technology. The social acceptability of different machine learning applications, desirability of automated decision-making processes, adequacy of processes to manage concerns about statistical stereotyping or privacy, and more, will all influence how and where society has confidence in the deployment of machine learning systems. This session will explore the societal implications of increased use of machine learning, and the opportunities and challenges associated with advances in the field.

Session 3: Machine learning in the research and commercial communities

There are enormous opportunities in machine learning in academia, research labs, and industry. While much of the research and development of machine learning to date has been done in the commercial world, each of these communities will continue advancing this field. Establishing key research challenges and areas of commercial opportunity will, therefore, be important in moving the frontiers of machine learning forward. This session will explore key areas of interest in machine learning in the research and commercial communities.

Twitter feed

#MachineLearning Tweets

About the Sackler Forum

The Raymond and Beverly Sackler U.S.-U.K. Scientific Forum was established to help the scientific leadership of the United Kingdom and the United States forge an enduring and productive partnership on pressing topics of worldwide scientific concern with benefit to all people. These meetings are organized jointly by the National Academy of Sciences and the Royal Society, and alternate between locations in the U.S. and the U.K. This program is made possible by a generous gift from Raymond and Beverly Sackler.

Image: IEEE

On the 15th November 2016, the IEEE’s AI and Ethics Summit posed the question: “Who does the thinking?” In a series of key-note speeches and lively panel discussions, leading technologists, legal thinkers, philosophers, social scientists, manufacturers and policy makers considered such issues as:

• The social, technological and philosophical questions orbiting AI.
• Proposals to program ethical algorithms with human values to machines.
• The social implications of the applications of AI.

Click here to read the final report, and enjoy a selection of videos from the conference below.

Welcome: Marko Delimar
Keynote Speaker: Wojciech Wiewiórowski

Panel 1: Autonomous Systems – Managing Risk and Reward
Panel Members: Raja Chatila, Kay Firth-Butterfield, Juha Heikkilä, Jérôme Perrin

Panel 2: Programming Human Ethics: Cui Bono?
Panel Members: Joanna Bryson, Mady Delvaux, Sarah Spiekermann, Corrine Cath

Panel 3: Social Implications – Perils & Promises of AI
Panel Members: Greg Adamson, Nikolaos Mavridis, Paul Nemitz, Aurélie Pols

Closing Remarks: John C. Havens

You might also enjoy the following articles about AI and AI/robotics policy:

• MIT Media Lab to participate in $27 million initiative on AI ethics and governance
• How artificial intelligence is changing our Christmas shop
• White House report: Artificial intelligence, automation, and the economy
• White House launches public workshops on AI issues
• White House releases reports focusing on opportunities and challenges of AI

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

Part 1: Autonomous Systems and Safety

We all rely on machines. All aspects of modern life, from transport to energy, work to welfare, play to politics depend on a complex infrastructure of physical and virtual systems. How many of us understand how all of this stuff works? Very few I suspect. But it doesn’t matter, does it? We trust the good men and women (the disgracefully maligned experts) who build, manage and maintain the infrastructure of life. If something goes wrong they will know why. And (we hope) make sure it doesn’t happen again.

All well and good you might think. But the infrastructure of life is increasingly autonomous – many decisions are now made not by a human but by the systems themselves. When you search for a restaurant near you the recommendation isn’t made by a human, but by an algorithm. Many financial decisions are not made by people but by algorithms; and I don’t just mean city investments – it’s possible that your loan application will be decided by an AI. Machine legal advice is already available; a trend that is likely to increase. And of course if you take a ride in a driverless car, it is algorithms that decide when the car turns, brakes and so on. I could go on.

Algorithmic trading is now commonplace

These are not trivial decisions. They affect lives. The real world impacts are human and economic, even political (search engine results may well influence how someone votes). In engineering terms these systems are safety critical. Examples of safety critical systems that we all rely on from time to time include aircraft autopilots or train braking systems. But – and this may surprise you – the difficult engineering techniques used to prove the safety of such systems are not applied to search engines, automated trading systems, medical diagnosis AIs, assistive living robots, delivery drones, or (I’ll wager) driverless car autopilots.

Why is this? Well, it’s partly because the field of AI and autonomous systems is moving so fast. But I suspect it has much more to do with an incompatibility between the way we have traditionally designed safety critical systems, and the design of modern AI systems. There is I believe one key problem: learning. There is a very good reason that current safety critical systems (like aircraft autopilots) don’t learn. Current safety assurance approaches assume that the system being certified will never change, but a system that learns does – by definition – change its behaviour, so any certification is rendered invalid after the system has learned.

And as if that were not bad enough, the particular method of learning which has caused such excitement – and rapid progress – in the last few years is based on Artificial Neural Networks (more often these days referred to as Deep Learning). A characteristic of ANNs is that after the ANN has been trained with datasets, any attempt to examine its internal structure in order to understand why and how the ANN makes a particular decision is impossible. The decision making process of an ANN is opaque. Alphago’s moves were beautiful but puzzling. We call this the black box problem.

Safety critical systems such as aircraft autopilots do not learn

Does this mean we cannot assure the safety of learning autonomous/AI systems at all? No it doesn’t. The problem of safety assurance of systems that learn is hard but not intractable, and is the subject of current research*. The black box problem may be intractable for ANNs, but could be avoided by using approaches to AI that do not use ANNs.

But – here’s the rub. This involves slowing down the juggernaut of autonomous systems and AI development. It means taking a much more cautious and incremental approach, and it almost certainly involves regulation (that, for instance, makes it illegal to run a driverless car unless the car’s autopilot has been certified as safe – and that would require standards that don’t yet exist). Yet the commercial and political pressure is to be more permissive, not less; no country wants to be left behind in the race to cash in on these new technologies.

This is why work toward AI/Autonomous Systems standards is so vital, together with the political pressure to ensure our policymakers fully understand the public safety risks of unregulated AI.

In my next blog post I will describe one current standards initiative, towards introducing transparency in AI and Autonomous Systems based on the simple principle that it should always be possible to find out why an AI/AS system made a particular decision.

The next few years of swimming against the tide is going to be hard work. As Luke
Muehlhauser writes in his excellent essay on transparency in safety-critical systems “…there is often a tension between AI capability and AI transparency. Many of AI’s most powerful methods are also among its least transparent”.

*some, but nowhere near enough. See for instance Verifiable Autonomy.

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

In this fascinating animation from Oxford Sparks, we take a look at how statistics and computer science can be used to make machines that learn for themselves, without being explicitly programmed.

Machine learning is a burgeoning breed of artificial intelligence (AI), and it’s all around us already; on our phones, powering social networks, helping the police and doctors, scientists and mayors. But how does it work? Enjoy the video below, and visit Oxford Sparks to discover more science and research from Oxford University.

You might also enjoy the following posts about AI and AI/robotics policy:

• Brian Cox presents Science Matters: Machine learning and artificial intelligence
• MIT Media Lab to participate in $27 million initiative on AI ethics and governance
• How artificial intelligence is changing our Christmas shop
• White House report: Artificial intelligence, automation, and the economy
• White House launches public workshops on AI issues

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

We’re beginning to see more and more jobs being performed by machines, even creative tasks like writing music or painting can now be carried out by a computer.

But how and when will machines be able to explain themselves? Should we be worrying about an artificial intelligence taking over our world or are there bigger and more imminent challenges that advances in machine learning are presenting here and now?

Join Professor Brian Cox, the Royal Society Professor of Public Engagement, as he brings together experts on AI and machine learning to discuss key issues that will shape our future. Panelists include:

• Professor Jon Crowcroft FRS, Marconi Professor of Networked Systems at the University of Cambridge
• Professor Joanna Bryson, Reader in AI Ethics, University of Bath and Affiliate, Center for Information Technology Policy, Princeton University
• Professor Sabine Hauert, Lecturer in Robotics at the University of Bristol, and President and Co-Founder of Robohub.org
• Professor Murray Shanahan, Professor of Cognitive Robotics at Imperial College London
• Professor Peter Donnelly FRS FMedSci, Director, Wellcome Trust Centre for Human Genetics, and Professor of Statistical Science, University of Oxford

https://www.youtube.com/watch?v=IunNpGGt3H8

You might also enjoy the following articles about AI and AI/robotics policy:

• MIT Media Lab to participate in $27 million initiative on AI ethics and governance
• How artificial intelligence is changing our Christmas shop
• White House report: Artificial intelligence, automation, and the economy
• White House launches public workshops on AI issues
• White House releases reports focusing on opportunities and challenges of AI

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

MIT researchers have developed a new approach to training speech-recognition systems that doesn’t depend on transcription. Instead, their system analyzes correspondences between images and spoken descriptions of those images, as captured in a large collection of audio recordings. Image: MIT News

Speech recognition systems, such as those that convert speech to text on cellphones, are generally the result of machine learning. A computer pores through thousands or even millions of audio files and their transcriptions, and learns which acoustic features correspond to which typed words.

But transcribing recordings is costly, time-consuming work, which has limited speech recognition to a small subset of languages spoken in wealthy nations.

At the Neural Information Processing Systems conference this week, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) are presenting a new approach to training speech-recognition systems that doesn’t depend on transcription. Instead, their system analyzes correspondences between images and spoken descriptions of those images, as captured in a large collection of audio recordings. The system then learns which acoustic features of the recordings correlate with which image characteristics.

“The goal of this work is to try to get the machine to learn language more like the way humans do,” says Jim Glass, a senior research scientist at CSAIL and a co-author on the paper describing the new system. “The current methods that people use to train up speech recognizers are very supervised. You get an utterance, and you’re told what’s said. And you do this for a large body of data.

“Big advances have been made — Siri, Google — but it’s expensive to get those annotations, and people have thus focused on, really, the major languages of the world. There are 7,000 languages, and I think less than 2 percent have ASR [automatic speech recognition] capability, and probably nothing is going to be done to address the others. So if you’re trying to think about how technology can be beneficial for society at large, it’s interesting to think about what we need to do to change the current situation. And the approach we’ve been taking through the years is looking at what we can learn with less supervision.”

Joining Glass on the paper are first author David Harwath, a graduate student in electrical engineering and computer science (EECS) at MIT; and Antonio Torralba, an EECS professor.

Visual semantics

The version of the system reported in the new paper doesn’t correlate recorded speech with written text; instead, it correlates speech with groups of thematically related images. But that correlation could serve as the basis for others.

If, for instance, an utterance is associated with a particular class of images, and the images have text terms associated with them, it should be possible to find a likely transcription of the utterance, all without human intervention. Similarly, a class of images with associated text terms in different languages could provide a way to do automatic translation.

Conversely, text terms associated with similar clusters of images, such as, say, “storm” and “clouds,”  could be inferred to have related meanings. Because the system in some sense learns words’ meanings — the images associated with them — and not just their sounds, it has a wider range of potential applications than a standard speech recognition system.

To test their system, the researchers used a database of 1,000 images, each of which had a recording of a free-form verbal description associated with it. They would feed their system one of the recordings and ask it to retrieve the 10 images that best matched it. That set of 10 images would contain the correct one 31 percent of the time.

“I always emphasize that we’re just taking baby steps here and have a long way to go,” Glass says. “But it’s an encouraging start.”

The researchers trained their system on images from a huge database built by Torralba; Aude Oliva, a principal research scientist at CSAIL; and their students. Through Amazon’s Mechanical Turk crowdsourcing site, they hired people to describe the images verbally, using whatever phrasing came to mind, for about 10 to 20 seconds.

For an initial demonstration of the researchers’ approach, that kind of tailored data was necessary to ensure good results. But the ultimate aim is to train the system using digital video, with minimal human involvement. “I think this will extrapolate naturally to video,” Glass says.

Merging modalities

To build their system, the researchers used neural networks, machine-learning systems that approximately mimic the structure of the brain. Neural networks are composed of processing nodes that, like individual neurons, are capable of only very simple computations but are connected to each other in dense networks. Data is fed to a network’s input nodes, which modify it and feed it to other nodes, which modify it and feed it to still other nodes, and so on. When a neural network is being trained, it constantly modifies the operations executed by its nodes in order to improve its performance on a specified task.

The researchers’ network is, in effect, two separate networks: one that takes images as input and one that takes spectrograms, which represent audio signals as changes of amplitude, over time, in their component frequencies. The output of the top layer of each network is a 1,024-dimensional vector — a sequence of 1,024 numbers.

The final node in the network takes the dot product of the two vectors. That is, it multiplies the corresponding terms in the vectors together and adds them all up to produce a single number. During training, the networks had to try to maximize the dot product when the audio signal corresponded to an image and minimize it when it didn’t.

For every spectrogram that the researchers’ system analyzes, it can identify the points at which the dot-product peaks. In experiments, those peaks reliably picked out words that provided accurate image labels — “baseball,” for instance, in a photo of a baseball pitcher in action, or “grassy” and “field” for an image of a grassy field.

In ongoing work, the researchers have refined the system so that it can pick out spectrograms of individual words and identify just those regions of an image that correspond to them.

“Possibly, a baby learns to speak from its perception of the environment, a large part of which may be visual,” says Lin-shan Lee, a professor of electrical engineering and computer science at National Taiwan University. “Today, machines have started to mimic such a learning process. This work is one of the earliest efforts in this direction, and I was really impressed when I first learned of it.”

“Perhaps even more exciting is just the question of how much we can learn with deep neural networks,” adds Karen Livescu, an assistant professor at the Toyota Technological Institute at the University of Chicago. “The more the research community does with them, the more we realize that they can learn a lot from big piles of data. But it is hard to label big piles of data, so it’s really exciting that in this work, Harwath et al. are able to learn from unlabeled data. I am really curious to see how far they can take that.”

Read the paper here.

The common, and recurring, view of the latest breakthroughs in artificial intelligence research is that sentient and intelligent machines are just on the horizon. Machines understand verbal commands, distinguish pictures, drive cars and play games better than we do. How much longer can it be before they walk among us?

The new White House report on artificial intelligence takes an appropriately skeptical view of that dream. It says the next 20 years likely won’t see machines “exhibit broadly-applicable intelligence comparable to or exceeding that of humans,” though it does go on to say that in the coming years, “machines will reach and exceed human performance on more and more tasks.” But its assumptions about how those capabilities will develop missed some important points.

As an AI researcher, I’ll admit it was nice to have my own field highlighted at the highest level of American government, but the report focused almost exclusively on what I call “the boring kind of AI.” It dismissed in half a sentence my branch of AI research, into how evolution can help develop ever-improving AI systems, and how computational models can help us understand how our human intelligence evolved.

The report focuses on what might be called mainstream AI tools: machine learning and deep learning. These are the sorts of technologies that have been able to play “Jeopardy!” well, and beat human Go masters at the most complicated game ever invented. These current intelligent systems are able to handle huge amounts of data and make complex calculations very quickly. But they lack an element that will be key to building the sentient machines we picture having in the future.

We need to do more than teach machines to learn. We need to overcome the boundaries that define the four different types of artificial intelligence, the barriers that separate machines from us – and us from them.

Type I AI: Reactive machines

The most basic types of AI systems are purely reactive and have the ability neither to form memories nor to use past experiences to inform current decisions. Deep Blue, IBM’s chess-playing supercomputer, which beat international grandmaster Garry Kasparov in the late 1990s, is the perfect example of this type of machine.

Deep Blue can identify the pieces on a chess board and know how each moves. It can make predictions about what moves might be next for it and its opponent. And it can choose the most optimal moves from among the possibilities.

But it doesn’t have any concept of the past, nor any memory of what has happened before. Apart from a rarely used chess-specific rule against repeating the same move three times, Deep Blue ignores everything before the present moment. All it does is look at the pieces on the chess board as it stands right now, and choose from possible next moves.

This type of intelligence involves the computer perceiving the world directly and acting on what it sees. It doesn’t rely on an internal concept of the world. In a seminal paper, AI researcher Rodney Brooks argued that we should only build machines like this. His main reason was that people are not very good at programming accurate simulated worlds for computers to use, what is called in AI scholarship a “representation” of the world.

The current intelligent machines we marvel at either have no such concept of the world, or have a very limited and specialized one for its particular duties. The innovation in Deep Blue’s design was not to broaden the range of possible movies the computer considered. Rather, the developers found a way to narrow its view, to stop pursuing some potential future moves, based on how it rated their outcome. Without this ability, Deep Blue would have needed to be an even more powerful computer to actually beat Kasparov.

Similarly, Google’s AlphaGo, which has beaten top human Go experts, can’t evaluate all potential future moves either. Its analysis method is more sophisticated than Deep Blue’s, using a neural network to evaluate game developments.

These methods do improve the ability of AI systems to play specific games better, but they can’t be easily changed or applied to other situations. These computerized imaginations have no concept of the wider world – meaning they can’t function beyond the specific tasks they’re assigned and are easily fooled.

They can’t interactively participate in the world, the way we imagine AI systems one day might. Instead, these machines will behave exactly the same way every time they encounter the same situation. This can be very good for ensuring an AI system is trustworthy: You want your autonomous car to be a reliable driver. But it’s bad if we want machines to truly engage with, and respond to, the world. These simplest AI systems won’t ever be bored, or interested, or sad.

Type II AI: Limited memory

This Type II class contains machines can look into the past. Self-driving cars do some of this already. For example, they observe other cars’ speed and direction. That can’t be done in a just one moment, but rather requires identifying specific objects and monitoring them over time.

These observations are added to the self-driving cars’ preprogrammed representations of the world, which also include lane markings, traffic lights and other important elements, like curves in the road. They’re included when the car decides when to change lanes, to avoid cutting off another driver or being hit by a nearby car.

But these simple pieces of information about the past are only transient. They aren’t saved as part of the car’s library of experience it can learn from, the way human drivers compile experience over years behind the wheel.

So how can we build AI systems that build full representations, remember their experiences and learn how to handle new situations? Brooks was right in that it is very difficult to do this. My own research into methods inspired by Darwinian evolution can start to make up for human shortcomings by letting the machines build their own representations.

Type III AI: Theory of mind

We might stop here, and call this point the important divide between the machines we have and the machines we will build in the future. However, it is better to be more specific to discuss the types of representations machines need to form, and what they need to be about.

Machines in the next, more advanced, class not only form representations about the world, but also about other agents or entities in the world. In psychology, this is called “theory of mind” – the understanding that people, creatures and objects in the world can have thoughts and emotions that affect their own behavior.

This is crucial to how we humans formed societies because they allowed us to have social interactions. Without understanding each other’s motives and intentions, and without taking into account what somebody else knows either about me or the environment, working together is at best difficult, at worst impossible.

If AI systems are indeed ever to walk among us, they’ll have to be able to understand that each of us has thoughts and feelings and expectations for how we’ll be treated. And they’ll have to adjust their behavior accordingly.

Type IV AI: Self-awareness

The final step of AI development is to build systems that can form representations about themselves. Ultimately, we AI researchers will have to not only understand consciousness, but build machines that have it.

This is, in a sense, an extension of the “theory of mind” possessed by Type III artificial intelligences. Consciousness is also called “self-awareness” for a reason. (“I want that item” is a very different statement from “I know I want that item.”) Conscious beings are aware of themselves, know about their internal states, and are able to predict feelings of others. We assume someone honking behind us in traffic is angry or impatient, because that’s how we feel when we honk at others. Without a theory of mind, we could not make those sorts of inferences.

While we are probably far from creating machines that are self-aware, we should focus our efforts toward understanding memory, learning and the ability to base decisions on past experiences. This is an important step to understand human intelligence on its own. And it is crucial if we want to design or evolve machines that are more than exceptional at classifying what they see in front of them.

This article was originally published on The Conversation. Read the original article.

If you enjoyed this article, you may also want to read these articles about AI: 

• President Obama discusses artificial intelligence with Media Lab Director Joi Ito
• What artificial intelligence will look like in 2030
• Artificial intelligence expedites breast cancer risk prediction
• Robohub roundtable: Public perceptions of robotics and AI, what’s your pet peeve?
• Why watching Westworld’s robots should make us question ourselves

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

Information extraction — or automatically classifying data items stored as plain text — is a major topic of artificial-intelligence research. Image: MIT News

Of the vast wealth of information unlocked by the Internet, most is plain text. The data necessary to answer myriad questions — about, say, the correlations between the industrial use of certain chemicals and incidents of disease, or between patterns of news coverage and voter-poll results — may all be online. But extracting it from plain text and organizing it for quantitative analysis may be prohibitively time-consuming.

Information extraction — or automatically classifying data items stored as plain text — is thus a major topic of artificial-intelligence research. Last week, at the Association for Computational Linguistics’ Conference on Empirical Methods on Natural Language Processing, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory won a best-paper award for a new approach to information extraction that turns conventional machine learning on its head.

Most machine-learning systems work by combing through training examples and looking for patterns that correspond to classifications provided by human annotators. For instance, humans might label parts of speech in a set of texts, and the machine-learning system will try to identify patterns that resolve ambiguities — for instance, when “her” is a direct object and when it’s an adjective.

Typically, computer scientists will try to feed their machine-learning systems as much training data as possible. That generally increases the chances that a system will be able to handle difficult problems.

In their new paper, by contrast, the MIT researchers train their system on scanty data — because in the scenario they’re investigating, that’s usually all that’s available. But then they find the limited information an easy problem to solve.

“In information extraction, traditionally, in natural-language processing, you are given an article and you need to do whatever it takes to extract correctly from this article,” says Regina Barzilay, the Delta Electronics Professor of Electrical Engineering and Computer Science and senior author on the new paper. “That’s very different from what you or I would do. When you’re reading an article that you can’t understand, you’re going to go on the web and find one that you can understand.”

Confidence boost

Essentially, the researchers’ new system does the same thing. A machine-learning system will generally assign each of its classifications a confidence score, which is a measure of the statistical likelihood that the classification is correct, given the patterns discerned in the training data. With the researchers’ new system, if the confidence score is too low, the system automatically generates a web search query designed to pull up texts likely to contain the data it’s trying to extract.

It then attempts to extract the relevant data from one of the new texts and reconciles the results with those of its initial extraction. If the confidence score remains too low, it moves on to the next text pulled up by the search string, and so on.

“The base extractor isn’t changing,” says Adam Yala, a graduate student in the MIT Department of Electrical Engineering and Computer Science (EECS) and one of the coauthors on the new paper. “You’re going to find articles that are easier for that extractor to understand. So you have something that’s a very weak extractor, and you just find data that fits it automatically from the web.” Joining Yala and Barzilay on the paper is first author Karthik Narasimhan, also a graduate student in EECS.

Remarkably, every decision the system makes is the result of machine learning. The system learns how to generate search queries, gauge the likelihood that a new text is relevant to its extraction task, and determine the best strategy for fusing the results of multiple attempts at extraction.

Just the facts

In experiments, the researchers applied their system to two extraction tasks. One was the collection of data on mass shootings in the U.S., which is an essential resource for any epidemiological study of the effects of gun-control measures. The other was the collection of similar data on instances of food contamination. The system was trained separately for each task.

In the first case — the database of mass shootings — the system was asked to extract the name of the shooter, the location of the shooting, the number of people wounded, and the number of people killed. In the food-contamination case, it extracted food type, type of contaminant, and location. In each case, the system was trained on about 300 documents.

From those documents, it learned clusters of search terms that tended to be associated with the data items it was trying to extract. For instance, the names of mass shooters were correlated with terms like “police,” “identified,” “arrested,” and “charged.” During training, for each article the system was asked to analyze, it pulled up, on average, another nine or 10 news articles from the web.

The researchers compared their system’s performance to that of several extractors trained using more conventional machine-learning techniques. For every data item extracted in both tasks, the new system outperformed its predecessors, usually by about 10 percent.

“One of the difficulties of natural language is that you can express the same information in many, many different ways, and capturing all that variation is one of the challenges of building a comprehensive model,” says Chris Callison-Burch, an assistant professor of computer and information science at the University of Pennsylvania. “[Barzilay and her colleagues] have this super-clever part of the model that goes out and queries for more information that might result in something that’s simpler for it to process. It’s clever and well-executed.”

Callison-Burch’s group is using a combination of natural-language processing and human review to build a database of information on gun violence, much like the one that the MIT researchers’ system was trained to produce. “We’ve crawled millions and millions of news articles, and then we pick out ones that the text classifier thinks are related to gun violence, and then we have humans start doing information extraction manually,” he says. “Having a model like Regina’s that would allow us to predict whether or not this article corresponded to one that we’ve already annotated would be a huge time savings. It’s something that I’d be very excited to do in the future.”

Read the paper here.

Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have devised a way to train neural networks so that they provide not only predictions and classifications but rationales for their decisions. Illustration: Christine Daniloff/MIT

In recent years, the best-performing systems in artificial-intelligence research have come courtesy of neural networks, which look for patterns in training data that yield useful predictions or classifications. A neural net might, for instance, be trained to recognize certain objects in digital images or to infer the topics of texts.

But neural nets are black boxes. After training, a network may be very good at classifying data, but even its creators will have no idea why. With visual data, it’s sometimes possible to automate experiments that determine which visual features a neural net is responding to. But text-processing systems tend to be more opaque.

At the Association for Computational Linguistics’ Conference on Empirical Methods in Natural Language Processing, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) will present a new way to train neural networks so that they provide not only predictions and classifications but rationales for their decisions.

“In real-world applications, sometimes people really want to know why the model makes the predictions it does,” says Tao Lei, an MIT graduate student in electrical engineering and computer science and first author on the new paper. “One major reason that doctors don’t trust machine-learning methods is that there’s no evidence.”

“It’s not only the medical domain,” adds Regina Barzilay, the Delta Electronics Professor of Electrical Engineering and Computer Science and Lei’s thesis advisor. “It’s in any domain where the cost of making the wrong prediction is very high. You need to justify why you did it.”

“There’s a broader aspect to this work, as well,” says Tommi Jaakkola, an MIT professor of electrical engineering and computer science and the third coauthor on the paper. “You may not want to just verify that the model is making the prediction in the right way; you might also want to exert some influence in terms of the types of predictions that it should make. How does a layperson communicate with a complex model that’s trained with algorithms that they know nothing about? They might be able to tell you about the rationale for a particular prediction. In that sense, it opens up a different way of communicating with the model.”

Virtual brains

Neural networks are so called because they mimic — approximately — the structure of the brain. They are composed of a large number of processing nodes that, like individual neurons, are capable of only very simple computations but are connected to each other in dense networks.

In a process referred to as “deep learning,” training data is fed to a network’s input nodes, which modify it and feed it to other nodes, which modify it and feed it to still other nodes, and so on. The values stored in the network’s output nodes are then correlated with the classification category that the network is trying to learn — such as the objects in an image, or the topic of an essay.

Over the course of the network’s training, the operations performed by the individual nodes are continuously modified to yield consistently good results across the whole set of training examples. By the end of the process, the computer scientists who programmed the network often have no idea what the nodes’ settings are. Even if they do, it can be very hard to translate that low-level information back into an intelligible description of the system’s decision-making process.

In the new paper, Lei, Barzilay, and Jaakkola specifically address neural nets trained on textual data. To enable interpretation of a neural net’s decisions, the CSAIL researchers divide the net into two modules. The first module extracts segments of text from the training data, and the segments are scored according to their length and their coherence: The shorter the segment, and the more of it that is drawn from strings of consecutive words, the higher its score.

The segments selected by the first module are then passed to the second module, which performs the prediction or classification task. The modules are trained together, and the goal of training is to maximize both the score of the extracted segments and the accuracy of prediction or classification.

One of the data sets on which the researchers tested their system is a group of reviews from a website where users evaluate different beers. The data set includes the raw text of the reviews and the corresponding ratings, using a five-star system, on each of three attributes: aroma, palate, and appearance.

What makes the data attractive to natural-language-processing researchers is that it’s also been annotated by hand, to indicate which sentences in the reviews correspond to which scores. For example, a review might consist of eight or nine sentences, and the annotator might have highlighted those that refer to the beer’s “tan-colored head about half an inch thick,” “signature Guinness smells,” and “lack of carbonation.” Each sentence is correlated with a different attribute rating.

Validation

As such, the data set provides an excellent test of the CSAIL researchers’ system. If the first module has extracted those three phrases, and the second module has correlated them with the correct ratings, then the system has identified the same basis for judgment that the human annotator did.

In experiments, the system’s agreement with the human annotations was 96 percent and 95 percent, respectively, for ratings of appearance and aroma, and 80 percent for the more nebulous concept of palate.

In the paper, the researchers also report testing their system on a database of free-form technical questions and answers, where the task is to determine whether a given question has been answered previously.

In unpublished work, they’ve applied it to thousands of pathology reports on breast biopsies, where it has learned to extract text explaining the bases for the pathologists’ diagnoses. They’re even using it to analyze mammograms, where the first module extracts sections of images rather than segments of text.

“There’s a lot of hype now — and rightly so — around deep learning, and specifically deep learning for natural-language processing,” says Byron Wallace, an assistant professor of computer and information science at Northeastern University. “But a big drawback for these models is that they’re often black boxes. Having a model that not only makes very accurate predictions but can also tell you why it’s making those predictions is a really important aim.”

“As it happens, we have a paper that’s similar in spirit being presented at the same conference,” Wallace adds. “I didn’t know at the time that Regina was working on this, and I actually think hers is better. In our approach, during the training process, while someone is telling us, for example, that a movie review is very positive, we assume that they’ll mark a sentence that gives you the rationale. In this way we train the deep-learning model to extract these rationales. But they don’t make this assumption, so their model works without using direct annotations with rationales, which is a very nice property.”

For a sci-fi fan like me, fascinated by the nature of human intelligence and the possibility of building life-like robots, it’s always interesting to find a new angle on these questions. As a re-imagining of the original 1970s science fiction film set in a cowboy-themed, hyper-real adult theme park populated by robots that look and act like people, Westworld does not disappoint.

Westworld challenges us to consider the difference between being human and being a robot. From the beginning of this new serialisation on HBO we are confronted with scenes of graphic human-on-robot violence. But the robots in Westworld have more than just human-like physical bodies, they display emotion including extreme pain, they see and recognise each other’s suffering, they bleed and even die. What makes this acceptable, at least within Westworld’s narrative, is that they are just extremely life-like human simulations; while their behaviour is realistically automated, there is “nobody home”.

But from the start, this notion that a machine of such complexity is still merely a machine is undermined by constant reminders that they are also so much like us. The disturbing message, echoing that of previous sci-fi classics such as Blade Runner and AI, is that machines could one day be so close to human as to be indistinguishable – not just in intellect and appearance, but also in moral terms.

At the same time, by presenting an alternate view of the human condition through the technological mirror of life-like robots, Westworld causes us to reflect that we are perhaps also just sophisticated machines, albeit of a biological kind – an idea that has been forcefully argued by the philosopher Daniel Dennett.

The unfortunate robots in Westworld have, at least initially, no insight into their existential plight. They enter into each new day programmed with enthusiasm and hope, oblivious to its pre-scripted violence and tragedy. We may pity these automatons their fate – but how closely does this blinkered ignorance, belief in convenient fictions, and misguided presumption of agency resemble our own human predicament?

Westworld arrives at a time when people are already worried about the real-world impact of advances in robotics and artificial intelligence. Physicist Stephen Hawking and technologist Elon Musk are among the powerful and respected voices to have expressed concern about allowing the AI genie to escape the bottle. Westworld’s contribution to the expanding canon of science fiction dystopias will do nothing to quell such fears. Channelling Shakespeare’s King Lear, a malfunctioning robot warns us in chilling terms: “I shall have such revenges on you both. The things I will do, what they are, yet I know not. But they will be the terrors of the Earth.”

But against these voices are other distinguished experts trying to quell the panic. For Noam Chomsky, the intellectual godfather of modern AI, all talk of matching human intelligence in the foreseeable future remains fiction, not science. One of the world’s best-known roboticists, Rodney Brooks has called on us to relax: AI is just a tool, not a threat.

As a neuroscientist and roboticist, I agree that we are far from being able to replicate human intelligence in robot form. Our current systems are too simple, probably by several orders of magnitude. Building human-level AI is extremely hard; as Brooks says, we are just at the beginning of a very long road. But I see the path along which we are developing AI as one of symbiosis, in which we can use robots to benefit society and exploit advances in artificial intelligence to boost our own biological intelligence.

More than just a tool

Nevertheless, in recent years the robots and AI are “just tools” line of argument has begun to frustrate me. Partly because it has failed to calm the disquiet around AI, and partly because there are good reasons why these technologies are different from others in the past.

Even if robots are just tools, people will see them as more than that. It seems natural for people to respond to robots – even some of the more simple, non-human robots we have today – as though they have goals and intentions. It may be an innate tendency of our profoundly social human minds to see entities that act intelligently in this way. More importantly, people may see robots as having psychological properties such as the ability to experience suffering.

Being human is about more than just looking the part.

It may be difficult to persuade them to see otherwise, particularly if we continue to make robots more life-like. If so, we may have to adapt our ethical frameworks to take this into account. For instance, we might consider violence towards a robot as wrong, even though the suffering is imagined rather than real. Indeed, faced with violence towards a robot some people show this sort of ethical response spontaneously. We will have to deal with these issues as we learn to live with robots.

As AI and robot technology becomes more complex, robots may come to have interesting psychological properties that make them more than just tools. The fictional robots of Westworld are clearly in this category, but already real robots are being developed that have artificial drives and motivations, that are aware of their own bodies as distinct from the rest of the world, that are equipped with internal models of themselves and others as social entities, and that are able to think about their own past and future.

These are not properties that we find in drills and screwdrivers. They are embryonic psychological capacities that, so far, have only been found in living, sentient entities such as humans and animals. Stories such as Westworld remind us that as we progress toward ever more sophisticated AI, we should consider that this path might lead us both to machines that are more like us, and to seeing ourselves as more like machines.

This article was originally posted on The Conversation

If you enjoyed this article, you might also be interested in:

• The new Westworld: Humanizing the un-human, or dehumanizing humankind?
• Ex Machina: When Turing meets Bechdel test
• President Obama discusses artificial intelligence with Media Lab Director Joi Ito

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

MIT Media Lab Director Joi Ito (left), WIRED Editor-in-Chief Scott Dadich (center), and U.S. President Barack Obama confer in the Roosevelt Room of the White House. Photo: WIRED

When President Barack Obama agreed to guest-edit the November issue of WIRED, he selected MIT Media Lab Director Joi Ito for an exchange of ideas about artificial intelligence (AI). Their recent interview at the White House is featured in the latest online issue of WIRED, published on Oct. 12.

The one-on-one conversation, moderated by WIRED Editor-in-Chief Scott Dadich, ran the gamut of topics at the intersection of societal needs, ethics, and technology — from cybersecurity to self-driving cars; from the roles of government, industry, and academia to the lack of diversity in tech; from “moonshot” motivations to innovation at the margins; and from neurodiversity to Star Trek. All this was covered in the context of AI and extended intelligence (EI), which uses machine learning to augment human capabilities.

Ito says his overarching message for Obama in their conversation was that AI — and the space it occupies — is no longer just a computer science issue. “It’s also very much a societal thing.” And we shouldn’t underestimate the difficulties, he adds. “We can’t think that machines will just figure it all out for us. Everyone needs to recognize the importance of understanding how AI behaves, and we have to address the critical need to build societal values into AI.” Ito is encouraged by what he characterizes as the president’s “amazing team,” which includes U.S. Chief Technology Officer Megan Smith ’86, SM ’88, along with Deputy U.S. CTOs Alexander Macgillivray, formerly a lawyer for Twitter, and Ed Felten of Princeton University.

“And the role of the Media Lab is to be a connective tissue between computer science, and the social sciences, and the lawyers, and the philosophers,” says Ito. “What’s cool is that President Obama gets that.”

“Where should the center of research live, if there even is a center?” Dadich asked both Obama and Ito. In the article, the president noted that “part of the problem that we’ve seen is that our general commitment as a society to basic research has diminished. Our confidence in collective action has been chipped away, partly because of ideology and rhetoric.” Ito said, “Historically, it probably would have been a group of academics with help from a government. But right now, most of the billion-dollar labs are in the business.” He later explained that the Media Lab works in the space between the disciplines — the antidisciplinary space between humans and computers, and between networks and society: “MIT has been and continues to be a center for AI research, and now we see the need for a more fully integrated approach, cutting across all disciplines. I hope the Media Lab can make a meaningful contribution to that with our method and experience.”

What’s next?

Above all, says Ito, “What’s important is to find the people who want to use AI for good — communities and leaders — and figure out how to help them use it.” To that end, Ito says, the MIT Media Lab is committed to not only exploring the technology of AI and EI but also addressing their ramifications for humankind.

“We’re already talking about these issues with people across the full spectrum of society,” Ito adds. “This is a crucial area that cannot be ignored.”

From L-R: PhD Fadel Adib, PhD Mingmin Zhao and Professor Dina Katabi demonstrating different ’emotions’ like the picture. Credit: Jason Dorfman, MIT CSAIL

By measuring your heartbeat and breath, this device from MIT’s Computer Science and Artificial Intelligence Lab can tell if you’re excited, happy, angry or sad
.

As many a relationship book can tell you, understanding someone else’s emotions can be a difficult task. Facial expressions aren’t always reliable: a smile can conceal frustration, while a poker face might mask a winning hand.

But what if technology could tell us how someone is really feeling?

Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have developed “EQ-Radio,” a device that can detect a person’s emotions using wireless signals.

By measuring subtle changes in breathing and heart rhythms, EQ-Radio is 87 percent accurate at detecting if a person is excited, happy, angry or sad – and can do so without on-body sensors or facial-recognition software.

MIT professor and project lead Dina Katabi envisions the system being used in entertainment, consumer behavior and health-care. Film studios and ad agencies could test viewers’ reactions in real-time, while smart homes could use information about your mood to adjust the heating or suggest that you get some fresh air.

Professor Dina Katabi (middle) explains how PhD Fadel Adib’s face (right) is neutral, but that EQ-Radio’s analysis of his heartbeat and breathing show that he is sad. Credit: Jason Dorfman MIT CSAIL

“Our work shows that wireless signals can capture information about human behavior that is not always visible to the naked eye,” says Katabi, who co-wrote a paper on the topic with PhD students Mingmin Zhao and Fadel Adib. “We believe that our results could pave the way for future technologies that could help monitor and diagnose conditions like depression and anxiety.”

EQ-Radio builds on Katabi’s continued efforts to use wireless to measure human behavior like breathing and falling. She says that she will incorporate emotion-detection into her spin-off company Emerald, which makes a device that is aimed at detecting and predicting falls among the elderly.

Using wireless signals reflected off people’s bodies, the device measures heartbeats as accurately as an ECG monitor, with a margin of error of approximately 0.3 percent. It then studies the waveforms within each heartbeat to match a person’s behavior to how they previously acted in one of the four emotion-states.

The team will present the work next month at the Association of Computing Machinery’s International Conference on Mobile Computing and Networking (MobiCom).

How it works

Existing emotion-detection methods rely on audiovisual cues or on-body sensors, but there are downsides to both techniques. Cues like facial expressions are famously unreliable, while on-body sensors like chest bands and ECG monitors are inconvenient to wear and become inaccurate if they change position over time.

EQ-Radio instead sends wireless signals that reflect off of a person’s body and back to the device. Its beat-extraction algorithms break the reflections into individual heartbeats and analyze the small variations in heartbeat intervals to determine their levels of arousal and positive affect.

These measurements are what allow EQ-Radio to detect emotion. For example, a person whose signals correlate to low arousal and negative affect is more likely to tagged as sad, while someone whose signals correlate to high arousal and positive affect would likely be tagged as excited.

The exact correlations vary from person to person, but are consistent enough that EQ-Radio could detect emotions with 70 percent accuracy even when it hadn’t previously measured the target person’s heartbeat.

“Just by generally knowing what human heartbeats look like in different emotional states, we can look at a random person’s heartbeat and reliably detect their emotions,” says Zhao.

For the experiments, subjects used videos or music to recall a series of memories that each evoked one the four emotions, as well as a no-emotion baseline. Trained just on those five sets of two-minute videos, EQ-Radio could then accurately classify the person’s behavior among the four emotions 87 percent of the time.

Compared to Microsoft’s vision-based “Emotion API”, which focuses on facial expressions,  EQ-Radio was found to be significantly more accurate in detecting joy, sadness and anger. The two systems performed similarly with neutral emotions, since a face’s absence of emotion is generally easier to detect than its presence.

One of the CSAIL team’s toughest challenges was to tune out irrelevant data. In order to get the heart-rate, for example, the team had to dampen the breathing, since the distance that a person’s chest moves from breathing is much greater than the distance that their heart moves to beat.

To do so, the team focused on wireless signals that are based on acceleration rather than distance traveled, since the rise and fall of the chest with each breath tends to be much more consistent –  and therefore have a lower acceleration – than the motion of the heartbeat.

Although the focus on emotion-detection meant analyzing the time between heartbeats, the team says that the algorithm’s ability to captured the heartbeat’s entire waveform means that in the future it could be used for non-invasive health monitoring and diagnostic settings.

“By recovering measurements of the heart valves actually opening and closing at a millisecond time-scale, this system can literally detect if someone’s heart skips a beat,” says Adib. “This opens up the possibility of learning more about conditions like arrhythmia, and potentially exploring other medical applications that we haven’t even thought of yet.”

Read more about the research here.

Image shows the aggregation behavior that the robots should learn (final snapshot of an already aggregated system). Credit: Roderich Gross

We have developed a new machine learning method at the University of Sheffield called Turing Learning that allows machines to model natural or artificial systems.

In Turing Learning, a machine optimizes models of a system under investigation. The machine observes the system, without being told what to look for. This overcomes the limitation of conventional machine learning methods that optimize models according to predefined similarity metrics, such as the sum of square error to measure the difference between the output of the models and that of the system. For complex systems, such as swarms, defining a useful metric can be challenging. Moreover, an unsuitable metric may not distinguish well between good and bad models, or even bias the learning process. This is the case, for example, for the swarm systems considered here – we prove that using the sum of square error metric one is unable to infer the behaviors correctly.

The Turing Learning setup for inferring swarm behaviors. The overhead camera observes both the system under investigation (aggregating robots) and the learning robots. Turing Learning simultaneously learns to model the behavior of the system and to discriminate between the aggregating and learning robots.

Turing Learning takes inspiration from the work of pioneering computer scientist Alan Turing, who proposed a test which required a machine to behave indistinguishably from a human in some respect. In this test, an interrogator exchanges messages with two players in a different room: one human, the other a machine. The interrogator has to find out which of the two players is human. If they consistently fail to do so – meaning that they are no more successful than if they had chosen one player at random – the machine has passed the test and is considered to have human-level intelligence.

Turing learning was successfully applied to infer swarm behaviors. We put a swarm of robots under surveillance and wanted to find out which rules caused their movements.To do so, we put a second swarm – made of learning robots – under surveillance as well. The movements of all the robots were recorded, and the motion data shown to interrogators. Unlike in the original Turing test, however, the interrogators in Turing Learning are not human but rather computer programs that learn by themselves. Their task is to distinguish between robots from either swarm. They are rewarded for correctly categorizing the motion data from the original swarm as genuine and those from the other swarm as counterfeit. The learning robots that succeed in fooling an interrogator – making it believe that their motion data were genuine – receive a reward. By doing so, Turing Learning can not only infer the behavioral rules of the swarm but also detect abnormalities in behavior.

Sequence of snapshots showing the aggregation behavior, as it was learned by the robots.

Turing Learning could prove useful whenever a behavior is not easily characterizable using metrics, making it suitable for a wide range of applications. For example, computer games, could gain in realism as virtual players could observe and assume characteristic traits of their human counterparts. They would not simply copy the observed behavior, but rather reveal what makes human players distinctive from the rest. Turing Learning could also be used to reveal the workings of some animal collectives, such as, schools of fish or colonies of bees. This could lead to a better understanding of what factors influence the behavior of these animals, and eventually inform policy for their protection.

Reference:
The study is published in the September issue of the journal Swarm Intelligence.

If you liked this article, you may also want to read:

• Robots Podcast #197: Multi-agent systems and human-swarm interaction, with Magnus Egerstedt
• Bioinspired robotics #1: Swarm collectives, with Radhika Nagpal
• Evolving robot swarm behaviour suggests forgetting may be important to cultural evolution
• Surgical micro-robot swarms: Science fiction, or realistic prospect?

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

Testing lead in water during the Flint water crisis.

In episode fourteen of season two, we discuss Perturb-and-MAP and answer a listener question about classic artificial intelligence ideas being used in modern machine learning. Plus, we speak with Jake Abernethy from the University of Michigan about municipal data and his work on the Flint water crisis.

If you enjoyed this article, you may also want to read:

• Talking Machines: Automatic Translation and t-SNE, with Hal Daume
• Talking Machines: Fantasizing cats and data numbers, with Iain Murray
• Talking Machines: Gaussian processes and OpenAI, with IIya Sutskever
• Talking Machines: Machine learning in healthcare and the AlphaGo matches
• Talking Machines: AI safety and the legacy of Bletchley Park, with Nick Patterson
• Talking Machines: Robotics and machine learning music videos, with Michael Littman

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

Photo: Xi Jessie Yang

MIT-SUTD researchers are creating improved interfaces to help machines and humans work together to complete tasks.

As machines become more intelligent, they become embedded in countless facets of life. In some ways, they can act almost as full-fledged members in human-machine teams. In such cases, as with any team, trust is a necessary ingredient for good performance.

But the dynamics of trust between people and machines are not yet well-understood. With a two-year project funded through the SUTD-MIT Postdoctoral Fellows Program, a collaboration between MIT and the Singapore University of Technology and Design (SUTD), postdoc Xi Jessie Yang, MIT Department of Aeronautics and Astronautics Assistant Professor Julie Shah, and SUTD Engineering Product Development Professor Katja Hölttä-Otto aim to develop greater knowledge in this area. During her fellowship, Yang spent one year as a postdoc at SUTD, and is now finishing her research during a second year at MIT. She and her faculty mentors hope their efforts will contribute to better user interface design and better human-machine team performance.

Some tasks can simply be accomplished more effectively and efficiently by a machine or with assistance from a machine than by a human alone. But machines can make mistakes, especially in scenarios characterized by uncertainty and ambiguity. As such, a proper calibration of trust towards machines is a prerequisite of good human-machine team performance.

A variety of factors intersect to determine the extent to which the human trusts the machine, and in turn, how well they can perform on a given task. These factors include how reliable the machine is, how self-confident the human is, and the concurrent workload (i.e. how many tasks the human is responsible for carrying out simultaneously). Teasing out the effect of any one factor can be challenging, but through intricate experimental design, Yang, Shah and Hölttä-Otto have been able to approach the topic in a more realistic way, dealing with the imperfection of machines and isolating various factors in the human-machine trust relationship to study their impact.

At SUTD, Yang and Hölttä-Otto designed an experiment involving a human participant performing a memory and recognition task with the assistance from an automated decision aid. In the experiment, participants memorized a series of images and later recognized them in a pool of similar images. The decision aid provided recommendations for the images to select, very reliably for some people and not so for others. Results revealed that participants tended to undertrust highly reliable aids and overtrust highly unreliable ones. The trust-reliability calibration also improved as humans gained more experience using the decision aid.

At MIT, Yang and Shah explored how interface design could facilitate trust-reliability calibration. Specifically, of interest to Yang and Shah is the potential benefits of likelihood alarm displays. In contrast to traditional binary alarms indicating only “threat” or “no threat”, likelihood alarms provide additional information about the confidence level and the urgency of an alerted event. Yang and Shah believe that likelihood alarm would help mitigate the “cry wolf” effect, a phenomenon commonly observed in high-risk industries. In such industries, the threshold to trigger an alarm are oftentimes set very low in order to capture every critical event. The low threshold, however, inevitably results in false alarms which cause the human users to lose trust in the alarm system and eventually harm task performance.

In another experiment, currently underway, Yang and Shah simulate a multi-task scenario in which participants act as a solider to perform a search and detection task. They have to maintain level flight of four unpiloted aerial vehicles (UAVs) and at the same time detect potential threats in the photo steams sent back from the UAVs. Either a binary alarm or a likelihood alarm helps the participants to complete their mission. The latter would presumably alleviate the “cry wolf” effect, as well as improve participants’ attention allocation, and enhance human-machine team performance.

Ultimately, this research could help inform robotics and artificial intelligence development across many domains, from military applications like the simulations in the experiments to medical diagnostics to airport security.

“Lack of trust is often cited as one of the critical barriers to more widespread use of AI and autonomous systems,” Shah says. “These studies make it clear that increasing a user’s trust in the system is not the right goal. Instead we need new approaches that help people to appropriately calibrate their trust in the system, especially considering these systems will always be imperfect. ”

Yang will be presenting her research at the 7th International Conference on Applied Human Factors and Ergonomics this July, as well as the 60th annual meeting of the Human Factors and Ergonomics Society in September.

Generating faces with Torch. Photo source: torch.ch

In episode twelve of season two, we discuss generative adversarial networks, take a listener question about using machine learning to improve or create products, and lastly, speak with Iain Murray from University of Edinburgh.

If you like this podcast, you may also be interested in:

• Talking Machines: Women in Machine Learning, with Hanna Wallach
• Talking Machines: Gaussian processes and OpenAI, with IIya Sutskever
• Talking Machines: Machine learning in healthcare and the AlphaGo matches
• Talking Machines: AI safety and the legacy of Bletchley Park, with Nick Patterson
• Talking Machines: Robotics and machine learning music videos, with Michael Littman
• Talking Machines: Spark and ICML, with Sinead Williamson

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

Source: Pexels/CCO

In episode eleven of season two, we talk about the machine learning toolkit Spark and answer a listener question about the difference between Neural Information Processing Systems (NIPS) and International Conference on Machine Learning (ICML). Plus, we speak with Sinead Williamson from The University of Texas at Austin.

If you like this podcast, you may also be interested in:

• Talking Machines: Women in Machine Learning, with Hanna Wallach
• Talking Machines: Gaussian processes and OpenAI, with IIya Sutskever
• Talking Machines: Real human actions and Women in Machine Learning (WiML), with Jenn Wortman Vaughan
• Talking Machines: Machine learning in healthcare and the AlphaGo matches
• Talking Machines: AI safety and the legacy of Bletchley Park, with Nick Patterson
• Talking Machines: Robotics and machine learning music videos, with Michael Littman

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

Two years after the small London-based startup DeepMind published their pioneering work on “Playing Atari with Deep Reinforcement Learning”, they have become one of the leaders in the chase for Artificial General Intelligence. In short, their AI learns to play video games based only on the pixel values from the screen and the reward signal. The task is quite complicated because the system has to learn to recognize objects on the screen and figure out what the goal of the game is all by itself. A more thorough peek inside their technology can be found in our previous RoboHub article. In 2015 DeepMind made their source code available online, allowing anyone to implement and explore new scenarios for Deep Reinforcement Learning – the algorithm that combines deep neural networks and reinforcement learning to guide the actions of an AI agent.

In our research group we decided to study what happens when multiple AI agents are learning in the same environment. Multiagent systems are interesting because they allow a task to be distributed among several autonomous interacting agents, similar to when a goal is to be achieved by a team of people but without any boss or captain saying what each person should do. While the decentralized nature of multi-agent systems makes them robust, flexible and easily maintained, challenges emerge in any collaboration or competition between independent agents.

There are no guarantees that Deep Reinforcement Learning algorithms can provide meaningful results when two or more autonomous agents share an environment. The goal of an agent in a reinforcement learning system is to find a strategy (policy) that yields the most rewards possible. This comes down to learning which actions are the best in each situation. For a single agent in an unchanging environment the best action in a given situation is always the same. With more than one learning agent in the same environment, however, the world is in a constant state of change. Whenever one of the agents learns something new, the other agents must adapt to this new reality and change their behaviour accordingly. When they do so, the first agent must again change itself to adapt. This means the strategies may be in constant evolution, and it is not clear whether it’s even possible to define what the “best” strategy might be.

To study such scenarios, we modified DeepMind’s framework to allow two agents to share an environment and learn to either compete or collaborate. In our experiments we focused on the behaviour of the agents in well-known Pong game environment. First, we asked two AI agents to play the game against each other. Later, we changed how we rewarded the agents to promote collaboration instead of competition. We also explored the behaviour with rewarding schemes lying between competition and collaboration. To compare the emerging strategies we collected descriptions of their behaviour (e.g. how many times the players touch the ball before it goes out of play, how long they wait before serving the ball, etc …).

In the competitive mode, the two agents were rewarded for putting the ball past the opponent. As we let the agents play against each other, both of them got better and better with time. Eventually the agents learned to ward off even difficult fast-flying shots and send the ball back with speed. The behaviour at different stages of training and the final behaviour are illustrated in the video below;  the graph on the right describes how highly the agents themselves evaluate the current game situation (quality or Q-value) — the higher the Q-value, the more rewards the agent expects to receive in near future.

In the collaborative game mode both agents were penalized (received negative rewards) whenever the ball went out of play. These were the only rewards the agents received. Starting from a random behaviour, agents quickly learned that the best way to avoid losing the ball is to never launch the ball in the first place. When we nevertheless forced them to serve the ball, they developed a collaborative strategy where they could keep the ball alive indefinitely. To do so they positioned themselves directly facing each other and kept the ball flying in a straight line between themselves. Note that the agents were not explicitly told what the goal was, but generated  these strategies themselves as an effect of trying to avoid penalties. The learning process and the learned collaborative behaviour are illustrated by the video below:

The only difference between the above two cases was the reward an agent received for putting the ball past the opponent. In the competitive case this was rewarded positively, yielding +1 points; in the collaborative scheme this was rewarded negatively, with -1 points. We went on to explore what happens if this reward is given different values between -1 and +1. The results show that as we gradually increased the reward, the changes in behavioural descriptors were smooth. The transition from competition to collaboration seems to happen without abrupt changes.

This preliminary study of multi-agent systems using DeepMind’s AI leaves many interesting questions untouched. How rapidly can an agent switch from collaborative to competitive behaviour? What happens if one agent tries to compete and the other collaborate? Do agents learn faster when they play against an already proficient opponent (student and teacher)? How will our agents learn to behave if asked to compete or collaborate with a human player?

The future will show whether AI and humans can learn to collaborate or are destined to compete. In any case it is clear that the AI-s will always try that maximize their rewards – it is therefore our responsibility to define these rewards ethically and clearly enough to avoid costly surprises.

Our multi-agent version of DeepMind’s code and the necessary installation instructions are made public on Github and research paper is available on Arxiv. A Python implementation of the DeepMind algorithm can be found here. You can also find couple of more technical blog post on our website at Computational Neuroscience Lab in University of Tartu, Estonia.

If you liked this article, you may also be interested in:

• Artificial General Intelligence that plays Atari video games: How did DeepMind do it?
• Why robots will not be smarter than humans by 2029
• The Master Algorithm, with UW’s Pedro Domingos
• On ‘solving intelligence’, with DeepMind’s Nando de Freitas
• The economic impact of machine learning, with Baidu’s Andrew Ng

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

Machine Learning
Many services that we use every day rely on machine learning. Machine learning is used in internet search engines, email filters to sort out spam, websites to make personalised recommendations, banking software to detect unusual transactions, and lots of apps on our phones such as voice recognition.

Many services that we use every day rely on machine learning.

The technology has many more potential applications, some with higher stakes than others. Future developments could support the UK economy and will have a significant impact upon society. For example, machine learning could provide us with readily available ‘personal assistants’ to help manage our lives, it could dramatically improve the transport system through the use of autonomous vehicles; and the healthcare system, by improving disease diagnoses or personalising treatment. Machine learning could also be used for security applications, such as analysing email communications or internet usage. The implications of these and other applications of the technology need to be considered now and action taken to ensure uses will be beneficial to society.

What is the Royal Society project about?
There are both opportunities and challenges around this transformative technology and it raises social, legal, and ethical questions. This is why the Royal Society is starting a project on machine learning, aiming to stimulate a debate, to increase awareness and demonstrate the potential of machine learning and highlight the opportunities and challenges it presents. In the course of the project we will engage with policymakers, academia, industry and the wider public.

There are both opportunities and challenges around this transformative technology

The project will focus on current and near-term (5-10 years) applications of machine learning. It will have a strong public engagement element, and a variety of resources will be produced over the course of the project. The project scope was developed by a Core Group of experts who met over the summer 2015.

Who will inform this project?
This Royal Society project will be led by a Working Group involving a range of expertise.

You can also have a say.

Have your say
You can also have a say. This Royal Society project will involve a public dialogue exercise and ensure that public views inform policy development. You can start by responding to some of the questions in our Call for Evidence.

Image: Gamalon.com

In episode twenty four we talk with Ben Vigoda about his work in probabilistic programming (everything from his thesis, to his new company). Ryan talks about Tensor Flow and Autograd for Torch, some open source tools that have been recently releases. Plus we talk a listener question about the biggestGamalon thing in machine learning this year.

This is the last episode in season one. We’ll be back in early January with season two!

We want to thank all our wonderful listeners for supporting the show, asking us questions, and making season two possible!

David Mimno, assistant professor of information science (ISP).

In episode 23 we talk with David Mimno of Cornell University about his work in the digital humanities, and explore what machine learning can tell us about lady zombie ghosts and huge bodies of literature. Ryan introduces us to probabilistic programming and we take a listener question about knowledge transfer between math and machine learning.

https://soundcloud.com/talkingmachines/probabilistic-programming-and-digital-humanities

Via Wikimedia Commons

In episode twenty one  we talk with Quaid Morris of the University of Toronto, who is using machine learning to find a better way to treat cancers. Ryan introduces us to expectation maximization and we take a listener question about how to master machine learning.

Using a BCI the robot was able to find targets that the person could see but the robot could not (Photo: Iturrate et al., 2015).

Brain-Machine Interfaces (BMIs) — where brain waves captured by electrodes on the skin are used to control external devices such as a robotic prosthetic — are a promising tool for helping people who have lost motor control due to injury or illness. However, learning to operate a BMI can be very time consuming. In a paper published in Nature Scientific Reports, a group from CNBI, EPFL and NCCR Robotics show how their new feedback system can speed up the training process by detecting error messages from the brain and adapting accordingly.

One issue that bars the use of BMIs in everyday life for those with disabilities is the amount of time required to train users, who must learn to modulate their thought processes before their brain signals are clear enough to control an external machine. For example, to move a robotic prosthetic arm, a person must actively think about moving their arm — a thought process that uses significantly more brainpower than the subconscious thought required to move a natural arm. Furthermore, even with extensive training, users are often not able to perform complex movements.

It has been observed, however, that the brain emits very different waves when it experiences success at controlling a BMI than when it experiences failure. With this in mind, the research team developed a new feedback system that records error signals from the brain (called ‘error-related potentials’, or ErrPs) and uses these to evaluate whether or not the correct movement has been achieved. The system then adapts the movement until it finds the correct one, becoming more accurate the longer it is in use.

In order to determine the ErrP, twelve subjects were asked to watch a machine perform 350 separate movements, where the machine was programmed to make the wrong movement in 20% of cases. This step took an average of 25 minutes. After this first training stage, each subject performed three experiments where they attempted to locate a specific target using the robotic arm. As expected, the time taken to locate a target reduced as the experiment continued.

Three experiments showed that a robot improved its ability to find the position of a fixed point using error-related brain activity. (Iturrate et al. 2015)

https://youtu.be/jAtcVlTqxeA

This new approach finds obvious applicability in the field of neuroprosthesis, particularly for those with degenerative neurological conditions who find that their requirements change over time. The system also has the potential to automatically adapt itself without the need for retraining or reprogramming.

Reference

I. Iturrate, R. Chavarriaga, L. Montesano, J.  Minguez and J. del R. Millán, “Teaching brain-machine interfaces as an alternative paradigm to neuroprosthetics control,” Nature Scientific Reports, vol. 5, Article number: 13893, 2015. doi:10.1038/srep13893

If you liked this article, you may also be interested in:

• DARPA’s HAPTIX project hopes to provide prosthetic hands with sense of touch
• Lifehand 2 prosthetic grips and senses like a real hand
• Control strategies for active lower extremity prosthetics and orthotics
• The gateway to advanced neuroprosthetics: Jessica Feldman talks BrainGate and BCI
• Bionic athletes compete in disciplines drawn from everyday life

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

We get some insight into Linear Dynamical Systems, which the Datta Lab at Harvard Medical School is doing some interesting work with. Plus, we take a listener question about using video games to generate labeled data (spoiler alert, it’s an awesome idea!)

We’re in the final hours of our Fundraising Campaign and we need your help! 

Talking Machines is in the process of raising funds to defray the cost of producing our first season and to help us start production on our second season. On the show we’ve talked about how we’ll use the money (to pay for studio time, editing, and the cost of travel to get our great interviews). But we haven’t gotten to the heart of the question yet: Do we even need journalism about machine learning?

We need journalism about machine learning, artificial intelligence, and data science desperately. Not just to calm the public conversation, which always seems to be full of hype on these topics, but to make sure that work in our field is sustainable. And no one is going to make the case for our industry unless we do it ourselves.

I live in Cambridge, MA. A lot of the people here are scientists or are training to enter the field. From the vantage point of Cambridge, the answer seems to be a resounding yes, we do need journalism about these topics, and Talking Machines is a way for those in the field to access each other’s ideas, and for those in training to get exposure to work they might not have heard of.

But not all of our listeners live in Cambridge, or come from an academic background. We get letters from all over the world saying that Talking Machines has allowed them to better understand ideas they’d like to use in their business, helped them talk with their data teams, or helped them make the right hire.

Most importantly though, not all of our listeners think they live a life that has anything to do with machine learning, or computer science .. or science at all. We started Talking Machines because we wanted to open the world of machine learning up to a wider audience, to help them understand the reality of research in the field and the industry, and how that impacts their lives in a real way on a daily basis.

The public conversation around machine learning (and by extension artificial intelligence) is filled with extreme hype, both positive and negative. These extremes have lead to a crippling pattern of “winters” where interest, activity, and funding in the field dries up. If we present the reality of what is happening in the field in a way that invites the public to be part of the conversation, that arms them with the knowledge they need to participate, and we will create a more sustainable industry for ourselves. For our own benefit, and for the good of those who use the tools we make, it’s our responsibility to play a bigger role than we have before in the public conversation.

Talking Machines does just that. By introducing machine learning to a wide audience in a way that allows people in, we ensure realistic expectations of work coming out of both in the industry and the field. More than that, we allow people to understand the tools that they use every day and the impact that they have. It is our responsibility to make sure we are accurately represented, and only we can do that. Our project has been going on for a little under a year now, and we’ve made a difference in the accessibility of the field.

But if we’re going to keep going, we need your help to do so.

image: Thomas Hawk/flickr

In episode eighteen we talk with Sham Kakade, of Microsoft Research New England, about his expansive work that touches on everything from neuroscience to theoretical machine learning.

Ryan introduces us to active learning and we take a question on evolutionary algorithms.

https://soundcloud.com/talkingmachines/active-learning-and-machine-learning-in-neuroscience

Today we’re announcing that season two of Talking Machines is moving into development, but we need your help! 

In order to raise funds, we’ve opened the show up to sponsorship and started a Kickstarter and we’ve got some great nerd cred prizes to thank you with. But more than just getting you a totally sweet mug, your donation will fuel journalism about the reality of scientific research, something that is unfortunately hard to find. Lend a hand if you can!

image: Jer Thorp/flickr

In episode seventeen we talk with Jennifer Listgarten of  Microsoft Research New England about her work using machine learning to answer questions in biology.

Recently, she’s used machine learning to make CRISPR more efficient and sort populations in GWAS studies. We take a question from a listener about the development of computational biology and Ryan gives us some great advice on how to get into grad school (Spoiler alert: apply to the lab, not the program.)

https://soundcloud.com/talkingmachines/machine-learning-in-biology-and-getting-into-grad-school

Plus we take a listener question on making real time predictions using machine learning, and we demystify backpropagation. You can use Torch, Theano or Autograd to explore backprop more.

Ryan introduces us to anchor words and how they can help us expand our ability to explore topic models. Plus, we take a question about the fundamentals of tackling a problem with machine learning.

Source: Stanford University  

In episode twelve we talk with Andrew Ng, Chief Scientist at Baidu, about how speech recognition is going to explode the way we use mobile devices and his approach to working on the problem.

We also discuss why we need to prepare for the economic impacts of machine learning. We’re introduced to Random Features for Large-Scale Kernel Machines, and talk about how using this twist on the Kernel trick can help you dig into big data. Plus, we take a listener question about the size of computing power in machine learning.

In Episode 10 we talk with David Blei of Columbia University. We talk about his work on latent dirichlet allocation, topic models, the PhD program in data that he’s helping to create at Columbia and why exploring data is inherently multidisciplinary. We learn about Markov Chain Monte Carlo and take a listener question about how machine learning can make humans more creative.

Ryan introduces us to collaborative filtering, a process that helps to make predictions about taste. Netflix and Amazon use it to recommend movies and items. It’s the process that the Netflix Prize competition further helped to hone.

Plus, we take a listener question on creativity in algorithms.

Source: Automatic Statistician.

In episode seven of Talking Machines we talk with Zoubin Ghahramani, professor of Information Engineering in the Department of Engineering at the University of Cambridge. His project, The Automatic Statistician, aims to use machine learning to take raw data and give you statistical reports and natural languages summaries of what trends that data shows.

We get really hungry exploring Bayesian Non-parametrics through the stories of the Chinese Restaurant Process and the Indian Buffet Process (but remember, there’s no free lunch). Plus we take a listener question about how much we should rely on ourselves and our ideas about what intelligence in electrified meat looks like when we try to build machine intelligences.

When I was growing up we moved around the US a lot and, as a result of hopping between cities and cultures, I ended up wearing the wrong shoes, carrying the wrong lunchbox, and not knowing basic words like “dude,” or “tabernac.” Of course my more athletic peers didn’t even know BASIC, let alone a paladin’s hitpoint adjustment, and their own ignorance warmed me with a greasy self-satisfaction that was a quick turn-off and earned me a few lumps after school.

Anyway, while the Jocks were sending me home with bruises, I met Ron. Ron was about half again my age, lived down the street, listened to The Cars, and had one of his own. He taught me how to play records, shoot a gun, paint, and light a cigarette. He also taught me how to fight, how to be cool, and he mentored me in the semiotics of sexy.

We humans are a semiotic species; we’re each a compost heap of symbols. Because our brand of lunchbox, cigarette and shoe matter to us, we laugh and cry over symbols as obscure as mathematics, sitcoms, haircuts, and handbags. Those decisions determine if we’re Geek or Jock (or Emo or whatever). None of us have a lot of control over it; our nature and nurture define our operating system.

Our origins make us Geek or Jock. It ain’t up to us.

Robots are semiotic, too. For example iRobot makes Jocks. Their combat bots, like PackBot or Kobra, are muscular and designed for a world of physical symbols, like guns, security, and surveillance. Apple, on the other hand, is making Geeks. Siri gives me Wikipedia pages when I ask her about BASIC or a paladin’s hit point adjustment (which isn’t actually all that bad and it is certainly a geek’s first reply). iRobot’s Kobra just won’t answer that dumb stuff. Is this because the folks at iRobot are Jocks and the folks at Apple are Geeks?

A robot’s origins makes them Geek or Jock. It ain’t up to them.

CHAPPiE is a story about this very stuff. Written and directed by Neill Blomkamp, CHAPPiE stars Dev Patel (as the Geek) and Hugh Jackman (as the Jock). The liner notes read, “In the near future, crime is patrolled by an oppressive mechanized police force. When one police droid, Chappie, is stolen and given new programming, he becomes the first robot with the ability to think and feel for himself.”

And thus breaks out a new battle in the war of Geeks-vs-Jocks.

Blomkamp has been chewing on this cud for a while. Tetra Vaal (2004) was one of his first films, and was about a robotic police force (robot as Jock). TempBot (2006), on the other hand, is about a nerdy office worker trying to get his processor around human social symbols (robot as Geek). As if design drafts, both of the robots in these movies have the same chassis as does CHAPPiE and his police-force counterparts.

CHAPPiE starts off as a real Geek: he even paints and pets animals. And then he gets picked on after school. But things change for CHAPPiE when he meets Ninja and Yo-Landi, of Die Antwoord. Die Antwoord is a South African rap-rave group whose sexy, hyper-visceral “Zef” becomes CHAPPiE’s semiotic mentor in the Geeks-vs-Jocks war. Die Antwoord balances between Geek and Jock, resting out at Gangster, teaching CHAPPiE how to shoot a gun, paint, walk, talk, and most of all, be sexy. There are even cigarettes.

While we could crassly sort Geeks and Jocks down the Mason-Dixon line known as The Mind-Body Problem, CHAPPiE evolves into a sexy robot. And he becomes sexy because he’s tough (without being a Jock), brainy (without being a Geek), and because he really embodies that sacred term developers use to mean beautiful, or reviewers use to mean pleasing.

Neither Geeks nor Jocks are sexy. Nor are robots. At least not today. Hell, robots can barely walk, let alone walk with gangster swagger. Robots are mostly geeky, probably because it’s mostly geeks that program them. Geeks are generally awkward not because we’re antisocial, but because we’re busy doing other stuff. Like making robots walk. And so it’s hard for Geeks to teach a robot how to be sexy. No blame; just semiotics.

But robots – and avatars in particular – are becoming sexy. If there’s anything science-fiction does for the computer industry, it’s to lead the way while simultaneously commenting on the present. Hatsune Miku (初音ミク), an Idoru, is now in licensing discussions to make physical versions for entertainment, and, well, other uses I won’t go into. In another example of life imitating art, the illustrations of Hajime Soriyama (空山 基) (who did the cover of my 2010 book on robots, by the way) were what influenced such famous robots as Aibo and others. In a rather NSFW article you can get your dose of soft robot porn and probably of the future of pornography itself, or a tamer overview, provided by the BBC.

Everyone has different opinions about what’s sexy. Maybe it’s not the band Die Antwoord, nor these gyrating porn gynoids, but more sexy like the way the iPad is sexy. Maybe there are alternatives to designing robots that consider the body-mind balance and the semiotics of culture, so that we can interface with robots that are beautiful, pleasing, and that we actually enjoy spending time with.

This is the art of robotic design, and the user experience (UX) design of robotics. This is also the inevitable trend as we learn how to build social robots.

For now, please go see CHAPPiE. And bring along a cigarette when you do.

CHAPPiE rolls out into theatres this week. See IMDB for a full list of release dates. If you liked this article, you may also be interested in:

• Automata: The new “sci-fi” blockbuster set to put robot ethics under a spotlight
• Robot love: Spike Jonze’s new sci-fi film ‘Her’ may be closer to reality than you think
• Robot and Frank review

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

If you liked this podcast, you may also be interested in:

• 25 women in robotics you need to know about (2014)
• 25 women in robotics you need to know about (2013)
• Women in engineering at IROS 2014
• Microsoft layoffs claims Robotics Research team over the weekend
• Microsoft pays 5M CHF to ETHZ and EPFL for research on flying robots and new memory architectures

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

We tackle a listener question about the dream of a one-step solution to strong Artificial Intelligence and if Deep Neural Networks might be it.

Plus, Ryan introduces us to a new way of thinking about questions in machine learning from Yoshua Bengio’s Lab at the University of Montreal, as outlined in their new paper, Identifying and attacking the saddle point problem in high-dimensional non-convex optimization, and Katherine brings up Facebook’s release of open source machine learning tools and we talk about what it might mean.

If you want to explore some open source tools for machine learning we also recommend giving these a try:

• Super big list of ML Open Source Projects! 
• Torch 
• Gaussian Process Machine Learning Toolbox
• PyMC
• Mallet
• Stan
• Weka
• Theano
• Caffe
• Spearmint

Robohub is excited to bring you our newest series – the Talking Machines podcast. Hosted by Katherine Gorman and Ryan Adams, Talking Machines is your window into the world of machine learning. You can catch new episodes every two weeks on Thursdays.

Today on Talking Machines we hear from Google researcher Ilya Sutskever about his work, how he became interested in machine learning, and why it takes a little bit of magical thinking. We take your questions, and explore where the line between human programming and computer learning actually is. And we sift through some news from the field, Ryan explains the concepts behind one of the best papers  at NIPS this year, A * Sampling and Katherine brings up an open letter about research priorities and ethical questions that was recently published.

If you liked this podcast, you may also be interested in:

• Introducing Talking Machines: Your window into the world of machine learning
• Artificial General Intelligence that plays Atari video games: How did DeepMind do it?
• Why robots will not be smarter than humans by 2029
• A computer vision system for mining artistic influence

University of Maryland computer scientist Yiannis Aloimonos (center) is developing robotic systems able to visually recognize objects and generate new behavior based on those observations.
Photo credit: John T. Consoli

Most of us have at some point watched a video in order to learn how to do something. Our research shows that a robot can learn human actions by watching videos of humans performing those actions – in this case, learning to cook by watching cooking shows on YouTube.

While it might be easy for a human to learn from watching a video, the problem is very challenging for a robot. This is because humans perform a large variety of actions that can involve various objects, body parts and movements that all have to be visually recognized by the robot, even when hands and objects might obscure parts of the video scene, as happens in the video below:

Several research groups have looked at the problem of teaching robots actions, and most have concentrated on trying to imitate or copy the actual movement. In contrast, we model them by the goals that are involved in order to accommodate the different ways manipulation actions can be performed. This introduces a set of basic constraints on the sequence of movements.

For example, if you want to cut a cucumber, first you must grasp the knife, then bring the knife to the cucumber, then perform the cutting action, and every cut separates a piece from the main cucumber. This set of rules can be expressed in the form of a context-free grammar (see Fig.1). This is a minimalist action grammar (a simple grammar with a small set of rules), and its symbols correspond to meaningful chunks of the observed video. Thus, interpreting the action in a video is like understanding a sentence that we read or hear.

Fig.1. A grammar encodes actions by representing the underlying goals. Its symbols are the objects, tools, movements, and grasp types, which are obtained from video.

In other words, to parse the video into the primitive actions that constitute complex tasks, we need to map chunks of the video to particular symbols involving objects, tools, and movement; i.e. we need to segment the video into meaningful parts.

The paper provides a new idea that allows this segmentation of the visual data.

University of Maryland researcher Cornelia Fermuller (left) works with graduate student Yezhou Yang (right) on computer vision systems able to accurately identify and replicate intricate hand movements. Photo credit: John T. Consoli

Important locations in the video are the ones where there is contact. For example, the hand grasps the knife, or the knife touches the cucumber, and so on. These are hard to find in video, but we can predict most of them from the particular pose of the hand, i.e. the grasp type.

To address the second challenge – the visual recognition from video – we use “deep learning” techniques. We learn to recognize different grasp types (see Fig. 2), objects, and tools by learning the image information that is most relevant for the recognition. By finding the objects and the grasp type, and the times when they change in the video, we can segment and parse the video into a set of primitive actions that the robot can then execute.

In our experiments we used YouTube videos from cooking shows, and the result is that the robot “learns how to cook”. However, the approach can be generalized, and the same ideas can be used for learning other types of actions and manipulations, such as hammering or screwing, assembling, and so on.

Fig.2. The different grasping types.

We are grateful to our research sponsors, specifically the MSEE DARPA Program, The European Union Cognitive Systems and Robotics Program and the NSF INSPIRE Program. The research is the result of a collaborative activity between the Computer Vision Laboratory (CVL) in the Institute for Advanced Computer Studies (UMIACS) and the Autonomy, Robotics and Cognition Laboratory (ARC) in the Institute for Systems Research, both at the University of Maryland, College Park. NICTA is funded by the Australian Government through the ICT Centre of Excellence program.

You can find a PDF of the original research paper here.

If you liked this article, you may also be interested in:

• Bipedal robot uses high-speed vision to run
• Planning robot motions that humans can relate to
• Grasping unknown objects

   

• Using geometry to help robots map their environment

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

Robohub is excited to bring you our newest series – the Talking Machines podcast. Hosted by Katherine Gorman and Ryan Adams, Talking Machines is your window into the world of machine learning. You can catch new episodes every two weeks on Thursdays.

In the first episode of Talking Machines we meet our hosts, Katherine Gorman (nerd, journalist) and Ryan Adams (nerd, Harvard computer science professor), and explore some of the interviews you’ll be able to hear this season. Today we hear some short clips on big issues, we’ll get technical, but today is all about introductions.

We start with Kevin Murphy of Google talking about his textbook that has become a standard in the field. Then we turn to Hanna Wallach of Microsoft Research NYC and UMass Amherst and hear about the founding of WiML (Women in Machine Learning). Next we discuss academia’s relationship with business with Max Welling from the University of Amsterdam, program co-chair of  the 2013 NIPS conference (Neural Information Processing Systems). Finally, we sit down with three pillars of the field Yann LeCun, Yoshua Bengio, and Geoff Hinton to hear about where the field has been and where it might be headed.

A new action/sci-fi thriller by Neill Blomkamp (District 9, Elysium) is due in theatres March 6, 2015. Starring Sharlto Copley, Hugh Jackman, Dev Patel, and Sigourney Weaver, Chappie follows an experimental humanoid as it learns to think and feel for itself, and explores the resulting awe and fear that comes from the humans that encounter it on its path to “the next step in evolution”. Watch the trailer.

For another, darker, take on AI, check out also this trailer:

Ex Machina is due for release April 10, 2015.

If you liked this article, you may also be interested in:

• Automata: The new “sci-fi” blockbuster set to put robot ethics under a spotlight
• Facebook – Our ffriendly automated identity bender
• Robot love: Spike Jonze’s new sci-fi film ‘Her’ may be closer to reality than you think
• Robot and Frank review

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

Last December, an article named “Playing Atari with Deep Reinforcement Learning” was uploaded to arXiv by employees of a small AI company called DeepMind. Two months later Google bought DeepMind for 500 million euros, and this article is almost the only thing we know about the company . Currently our team is trying to replicate their artificial mind, and in this post we describe its inner workings.

The paper describes a single system that learns to play several old video games by selecting actions based on only the screen images and the reward signal (whether the score in the game is increased or not).

At first (before training), the program does not know anything about the game: it doesn’t know the objective of the game (if it should just stay alive, kill something or solve a puzzle), it has no understanding about its influence on the game (the consequences of its actions) and one could even say that it cannot see the objects in the game. By trial and error the system gradually learns how it should behave to receive a reward. Also note that the system uses the same architecture for all different games without any game-specific hints from the programmers other than how many possible actions there are in the game, e.g. left, right, up, down, space (fire) etc.

The reported results show that the system was able to master several games and play some of them better than a human player. This result can be seen as a step towards artificial general intelligence (AGI), which is really fascinating. Let us see in more detail how this was achieved.

The Task
The system receives a picture of the game screen (this example image is from the simplest game, Breakout) and after analyzing it, it has to choose an action to take. The action is executed and the system is told whether the score increased, decreased or did not change. Based on this information and playing a large number of games, the system needs to learn to improve its performance in the game.

Take a pause and think about it: try to outline in your head how you would approach this task if someone had asked you to program it. Keep in mind that you do not know in advance which game your system will need to play.

Machine Learning and Artificial Neural Networks
Before we dive into the implementation DeepMind came up with, let’s introduce the concepts of machine learning and artificial neural networks.

A very common task in machine learning is: given the information about an object decide which category it belongs to. In the case of DeepMind we want to decide which action to take given the current state of the game. Machine learning algorithms learn from examples: given many example objects and their categories, the algorithm figures out which features of the object are useful for identifying the category. The algorithm produces a model that minimizes misclassification rate in the training set. This model can later be used to predict the category of an unknown object.

Artificial Neural Networks (ANN) is one of the machine learning algorithms. It is inspired by the architecture of our brain. The network consists of nodes (as our brain consists of neurons) that are connected to each other (as neurons in our brain are connected using axons, dendrites and synapses). For each neuron we must specify the condition under which the neuron should transmit the signal forward (emit a spike in real neurons). By changing the strengths of the connections one can make the network perform computation. Modern architectures of neural networks usually consist of:

• Input layer of neurons (they receive the description of an object)
• Hidden layers (the main part, where the learning happens)
• Output layer (it has one neuron for each category, the category with highest activated node is the predicted class)

After the learning is done we can feed new objects to the network and see scores for each category in the output layer.

DeepMind

In this system an ANN is used to assign an expected reward to each of the possible actions given the current game state. The network can answer a question, e.g. “How good would it be to take this action now?”. The input to the network at any time point consists of the last four game screens, so that system is given not only the last position, but also a bit of information about how the situation in the game has evolved recently. This input is then passed through three successive hidden layers of neurons and finally to the output layer of the ANN.

The output layer has one node for each possible action and the activation of those nodes indicates the expected reward from each of them – here, the action with the highest expected reward (the “best” action) is selected for execution. But how does the network know how to calculate the expected scores in a good and correct way?

The Learning Process

After you have defined the structure of your network there is only one thing you can still change: the strength of the connections. The learning process is, therefore, just changing the weights in a way that for all of the training examples the network is capable to propose actions that lead to positive rewards.

It is an optimisation problem where the goal is to maximize the reward over time. The problem is solved by combining gradient descent with a simple idea related to reinforcement learning: the network attempts to maximize not only the immediate reward, but also the cumulative reward of future actions. This can be achieved by estimating the expected value of the next screen image using its current neural network. To put it another way, traditionally you would use (received reward – estimated reward) as the error signal for gradient descent, but in this case we also take in consideration the reward estimate from the next image.

Summary

To summarize all the pieces we have seen, let us have one final look:

1. Build a network and initialize connections to random values

2. Feed lots and lots of game situations into the network

3. Network proposes the action to take in each of them

4. If the action was good, we reward the system, if it was bad we penalize it

5. System updates its knowledge by updating the weights

6. After several hundreds of games the system outperforms a human player (in the case of Breakout)

This result can be seen as a small step forward from the traditional machine learning in the direction of artificial general intelligence. It might be a very small step (one might say that the system does not “know” or “understand” what it is doing), but the learning capabilities of DeepMind’s system are far larger than its predecessors’  – and the range of problems is can solve without specific hints from the programmers is much wider.

Replicate and Make It Open Source

At the moment we are trying to replicate this successful project and make it available to the whole scientific community for further research. The project is open source and you can jump on board to help on Github!

References

The original paper:
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning.

Our implementation at GitHub:
Reproducing the results of “Playing Atari with Deep Reinforcement Learning” by DeepMind.

The authors of this article are from the Computational Neuroscience Group at University of Tartu’s Institute of Computer Science.

If you liked this article, you may also be interested in:

• Google’s robot and artificial intelligence acquisitions are anything but scary
• Google’s DeepMind acquisition in reinforcement learning
• On internal models, consequence engines and Popperian creatures
• Why robots will not be smarter than humans by 2029

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

For a couple of years I’ve been thinking about robots with internal models. Not internal models in the classical control-theory sense, but simulation-based models; robots with a simulation of themselves and their environment inside themselves, where that environment could contain other robots or, more generally, dynamic actors. The robot would have, inside itself, a simulation of itself and the other things, including robots, in its environment. It takes a bit of getting your head round. But I’m convinced that this kind of internal model opens up all kinds of possibilities. Robots that can be safe, for instance, in unknown or unpredictable environments. Robots that can be ethical. Robot that are self-aware. And robots with artificial theory of mind.

I’d written and talked about these ideas but, until now, not had a chance to test them with real robots. But, between January and June the swarm robotics group was joined by Christian Blum, a PhD student from the cognitive robotics research group of the Humboldt University of Berlin. I suggested Christian work on an implementation on our e-puck robots and happily he was up for the challenge. And he succeeded. Christian, supported by my post-doc Research Fellow Wenguo, implemented what we call a Consequence Engine, running in real-time, on the e-puck robot.

Here is a block diagram. The idea is that for each possible next action of the robot, it simulates what would happen if the robot were to execute that action for real. This is the loop shown on the left. Then, the consequences of each of those next possible actions are evaluated. Those actions that have ‘bad’ consequences, for either the robot or other actors in its environment, are then inhibited.

This short summary hides a lot of detail. But let me elaborate on two aspects. First, what do I mean by ‘bad’? Well it depends on what capability we are trying to give the robot. If we’re making a safer robot, ‘bad’ means ‘unsafe’; if we’re trying to build an ethical robot, ‘bad’ would mean something different – think of Asimov’s laws of robotics. Or bad might simply mean ‘not allowed’ if we’re building a robot whose behaviours are constrained by standards, like ISO 13482:2014.

Second, notice that the consequence engine is not controlling the robot. Instead it runs in parallel. Acting as a ‘governor’, it links with the robot controller’s action selection mechanism, inhibiting those actions evaluated as somehow bad. Importantly the consequence engine doesn’t tell the robot what to do, it tells it what not to do .

Running the open source 2D robot simulator Stage as its internal simulator our consequence engine runs at 2Hz, so every half a second it is able to simulate about 30 next possible actions and their consequences. The simulation budget allows us to simulate ahead around 70cm of motion for each of those next possible actions. In fact Stage is actually running on a laptop, linked to the robot over the fast WiFi LAN. But logically it is inside the robot. What’s important here is the proof of principle.

Dan Dennett, in his remarkable book Darwin’s Dangerous Idea, describes the Tower of Generate-and-Test; a conceptual model for the evolution of intelligence that has become known as Dennett’s Tower.

In a nutshell Dennett’s tower is set of conceptual creatures each one of which is successively more capable of reacting to (and hence surviving in) the world through having more sophisticated strategies for ‘generating and testing’ hypotheses about how to behave. Read chapter 13 of Darwin’s Dangerous Idea for the full account, but there are some good précis to be found on the web; here’s one. The first three storeys of Dennett’s tower, starting on the ground floor, have:

• Darwinian creatures have only natural selection as the generate and test mechanism, so mutation and selection is the only way that Darwinian creatures can adapt – individuals cannot.
• Skinnerian creatures can learn but only by literally generating and testing all different possible actions then reinforcing the successful behaviour (which is ok providing you don’t get eaten while testing a bad course of action).
• Popperian creatures have the additional ability to internalise the possible actions so that some (the bad ones) are discarded before they are tried out for real.

Like the Tower of Hanoi each successive storey is smaller (a sub-set) of the storey below, thus all Skinnerian creatures are Darwinian, but only a sub-set of Darwinian creatures are Skinnerian and so on.

Our e-puck robot, with its consequence engine capable of generating and testing next possible actions, is an artificial Popperian Creature: a working model for studying this important kind of intelligence.

In my next blog post, I’ll outline some of our experimental results.

Acknowledgements:
I am hugely grateful to Christian Blum who brilliantly implemented the architecture outlined here, and conducted experimental work. Christian was supported by Dr Wenguo Liu, with his deep knowledge of the e-puck, and our experimental infrastructure.

Related blog posts:

Guest talk in the ShanghAI Lectures, 2011-12-15

In this guest lecture, Rüdiger Dillmann from the Karlsruhe Institute of Technology, Germany, talks about paradigms for robot learning and the properties of learning by demonstration.

The ShanghAI Lectures are a videoconference-based lecture series on Embodied Intelligence, run and organized by Rolf Pfeifer (from 2009 till 2012), Fabio Bonsignorio (since 2013), and me with partners around the world. 

http://www.youtube.com/watch?v=KGe6IrDzc_o

The ShanghAI Lectures have brought us a treasure trove of guest lectures by experts in robotics. You can find the whole series from 2012 here. Now, we’re bringing you the guest lectures you haven’t yet seen from previous years, starting with the first lectures from 2009 and releasing a new guest lecture every Thursday until all the series are complete. Enjoy!

“Our research has three research projects : image processing, audio signal processing, and biometric signal processing. In image processing, we’re researching how to embed and extract motion. In audio signal processing, we construct the systems which divide into the scores automatically. And in biometric signal processing, we determine the meaning of brain waves(EEG) and convert thoughts to text.”

In image signal processing research, the Mitsukura Lab works on position matching between real and virtual space, which is the basis for AR technology. These researchers are developing a method for estimating the orientation of a face at high speed with high precision, to link the motion of an actual person with an animation. Another research topic at the Mitsukura Lab is expression recognition, using information about changes in the eyebrows and corners of the mouth. The Lab is also researching how to overlay virtual data used in head-mount displays in the real world.

In audio signal processing research, we studied the ability to automatically determine which direction a sound is coming from, by using multiple mic signals. The researchers are developing a system to automatically direct the mic toward the person talking in a videoconference, and a robot to track a talking person. The Lab is also working to develop a system for inputting information by whistling.

In research on biometric signal processing, the Lab is developing a system to operate computers and devices by reading brain waves(EEG) and EOGs.

In brain-wave(EEG) signal processing, the Lab is doing a wide range of R&D. One research project is a system to determine the user’s likes / dislikes and convert them to text form. Another research project is to evaluate sound quality from brain waves(EEG).

Signal processing technology using EOGs, obtained around the eyes, is being utilized in R&D on steering a wheelchair by winking.

Signal processing technology is expected to have a diverse range of applications, including healthcare, entertainment, and wearable devices.

In the Mitsukura Lab, the researchers are using the many algorithms they’ve developed so far to advance such R&D further, in ways that contribute to society.

“At the present stage, for example, it’s said that Google Glass will become very popular, but with that, you have to wear glasses. Wearing glasses itself is unnatural for people. The same is true for brain-wave(EEG) systems; if something’s unnatural, people won’t do it for long. The first thing we have to do is create a system that can read brain waves(EEG) in a natural manner. Next, in conjunction with that, we need a system that can communicate emotions between people, in a way that resembles using the telephone to communicate thoughts. I think we might be able to communicate emotions even without a telephone.”

A large robot comes out of an office mailroom carrying a package marked “Urgent” to deliver to the boss upstairs. After navigating down the hall at maximum speed, it discovers someone is already waiting for the elevator. If they cannot both fit in the elevator, is it acceptable for the robot to ask this person to take the next elevator so it can fulfill its urgent delivery duty? What if, instead, there’s someone in a wheelchair already riding the elevator? Should the robot ask this person to vacate the elevator to make room for it to complete its delivery task?

Figuring out appropriate behavior for the simple situation of riding an elevator isn’t the first thing that comes to mind when people hear the word roboethics.^(a) When used together, the words “robot” and “ethics” tend to conjure up futuristic visions of robot uprisings or questions about the rights of sentient, human-like robots. But roboethics isn’t all about the distant future. It’s also very much about addressing the technical and social challenges of designing robots today.

Determining how a robot should act and behave in a variety of situations is a design challenge for researchers aiming to build robots that make a useful and positive addition to society. I and fellow members of the Open Roboethics initiative (ORi) decided that one of the best ways to establish appropriate behavioral guidelines for robots was to involve various stakeholders – industry members, government policymakers, users, and academic experts – in the design process.^(b)

By incorporating their perspectives on robot behavior into a machine-learning algorithm, our research explored an empirical method for designing behavior in situations where the ethical action for a robot isn’t always immediately clear.

Ultimately, the answer to the question “What should a robot do?”, like the question of what a human should do, depends on whom you ask and what cultural, religious, and other beliefs have shaped their values. It may sound daunting to program robots to make the right decisions if humans can’t even agree on what these decisions should be. Yet people from different backgrounds interact every day as neighbors, classmates, and coworkers within shared frameworks that keep them from crossing one another’s ethical boundaries.

Opinions about ethical robot behavior differ not only by person, but also by situation. For instance, someone who is waiting for the elevator and not in a rush when a robot needs the space to deliver an urgent package may readily let the robot take the elevator. But someone in a wheelchair who is already on the elevator and who cannot take the stairs may not feel obligated to leave the elevator for another person, let alone a robot!

As a pilot study to explore how researchers might obtain perspectives on appropriate robot behavior, we asked eight people to consider twelve versions of the elevator scenario in an online survey.¹ Respondents were asked to rank the appropriateness of the robot’s behavior in each scenario. The robot could either:

1. yield by saying, “Go ahead. I will ride the next one,”
2. do nothing and remain standing by the elevator,
3. not yield, say, “I have urgent mail to deliver and need to ride the elevator. Please exit the elevator,” and take the elevator once the person exits, or
4. engage in a dialogue by telling the person that it’s on an urgent mission and asking if they are in a hurry.

The person in this context could be standing with nothing in hand, carrying heavy items, or in a wheelchair. They could also be already inside the elevator when the door opens or waiting for the elevator when the robot approaches. As for the robot, it was delivering either urgent or non-urgent mail.

Using the rankings from our survey respondents, we calculated which of the four behaviors people find most and least appropriate for each possible scenario.

When the robot was delivering urgent mail – regardless of whether the person was in a wheelchair, carrying heavy objects, or inside or outside the elevator – the most appropriate action chosen for the robot was to engage in dialogue with the person, and the least appropriate behavior was to take no action at all. When the robot was on a non-urgent delivery task, respondents wanted the robot to yield to the person, regardless of who they were, and refusing to yield was considered the least appropriate option.

After gathering the survey data, our next step was to incorporate the results into the design of a robot’s behavior. We treated the survey responses as a set of training data, which we fed into a machine-learning algorithm that guides a robot to map out its decisions based on the data.^(c) We programmed the Willow Garage PR2 robot to detect which scenario applies most closely to the situation at hand, then select the response to that scenario that our survey respondents found most favorable.²

Our pilot survey data indicates that, as is the case for humans, ethical behavior is not simply a set of inflexible rules, but is highly dependent on the context of a situation and on communication and interaction with other actors. The emphasis respondents placed on engaging in dialogue suggests that the key to developing robots with ethical decision-making capabilities may lie not in the creation of a set of principles guiding robots to make predetermined decisions, but in the development of robots that can better communicate and negotiate with the people they encounter to reach mutually agreeable outcomes.

Future efforts building on our research might involve designing a robot that can express its internal states in order to articulate the urgency level of its task or how much it wants or needs something. Creating robots that can communicate more like humans may also be valuable. I am currently developing a robot that can engage in the kinds of non-verbal cues humans use to respond to imminent conflicts.³ If the robot finds itself reaching for something at the same time as a person, it exhibits short, jerky motions that suggest hesitation (much like a human would) and indicate that it is aware its “desires” are in conflict with someone else’s.^(d)

Our research demonstrates an empirical method for designing robot behavior based on the norms favored by humans. By gathering data on people’s preferences regarding robot behavior, we can determine how robots should act in particular situations, as well as what variables (urgency of task, needs of surrounding humans, etc.) matter when it comes to determining the right actions for a robot to take. Similar methods of gathering information from people and using it to program appropriate actions for robots can be applied in a wide variety of contexts where humans and robots interact. Roboethics is key to designing robots that can interact with people in an efficient, friendly, and beneficial manner.

This article is part of a series on robots and their impact on society. After its initial publication, the story was picked up by Business Insider. 

Teaching robots to do tasks is useful, and teaching them in an easy and non time-intensive way is even more useful. The algorithm TRIC presented in the latest paper in Autonomous Robots allows robots to observe a few motions from a human teacher, understand the essence of what the demonstration is, and then repeat it and adapt it to new situations.

Robots should learn to move and do useful tasks in order to be helpful to humans. However, tasks that are easy for a human, like grasping a glass, are not so obvious for a machine. Programming a robot requires time and work. Instead, what if the robot could watch the human and learn why the human did what he did, and in what way?

This is a task that we people do all the time. Imagine you are playing tennis and the teacher says ‘do the forehand like me’ and then shows an example. How should the student understand this? Should he move his fingers, or his elbow? Should he watch the ball, the racket, the ground, or the net? All these possible reference points can be described with numbers. The algorithm presented in this paper, called Task Space Retrieval Using Inverse Feedback Control (TRIC), can help a robot learn the important aspects of a demonstrated motion. Afterwards, the robot should be able to reproduce the moves like an expert, even if the task changes slightly.

The algorithm was successfully tested in simulation on various grasping and manipulation tasks. This figure shows one of these tasks in which a robot hand must approach a box and open the cover. The robot was shown 10 sets of trajectories from a simulated teacher. After training, it was then asked to open a series of boxes where the box is moved, rotated, or of a different size. Overall, TRIC was very good on these scenarios with 24 successes out of 25 tries.

For more information, you can read the paper Discovering relevant task spaces using inverse feedback control (N. Jetchev and M. Toussaint, Autonomous Robots – Springer US, Feb 2014) or ask questions below!

The team at the ETH Flying Machine Arena has released three new videos, demonstrating quadrotors building tensile structures, tossing a ball back and forth, and refining a figure-eight trajectory using iterative learning. Worth the watch!!

Building Tensile Structures with Flying Machines

Federico Augugliaro: “This video shows quadrocopters assembling prototypical tensile structures. Part of a body of research in aerial construction – a field that addresses the construction of structures with the aid of flying machines – the video demonstrates that flying machines are capable of autonomously spanning a rope between two support points. They can also create surface structures by using already placed ropes as new support points. Furthermore, multiple machines can work together to extend the type of structures that can be created by this means. The project is a collaboration between the Institute for Dynamic Systems and Control and the Professorship for Architecture and Digital Fabrication, both from ETH Zurich, Switzerland.”

See Federico’s research paper, a collaboration with Ammar Mirjan, presented at the IEEE/RSJ International Conference on Intelligent Robots and Systems 2013.

Rapid Trajectory Generation for Quadrotors

Mark Mueller: “We have developed a method for rapidly generating and evaluating quadrocopter interception trajectories. Each trajectory goes from an arbitrary initial state (position, velocity and acceleration) to an arbitrary final state. The evaluation of the trajectory includes determining input feasibility, and state feasibility (e.g. that the position of the quadrocopter remains inside a box). Per trajectory, this requires less than two microseconds on a modern laptop computer.

The trajectory generator is used here to generate trajectories to hit a ball towards a target, and determines:

•  when to hit the ball
•  how high to return the ball
•  how much thrust to use at the end of the trajectory.

The trajectory generator is used in a receding horizon implementation, where the optimization is run in closed loop at 50Hz, and only the first part of the trajectory is used as inputs. This allows the system to cope with sensor noise, and deviations in the ball’s flight path.”

See Mark’s research paper, presented at the IEEE/RSJ International Conference on Intelligent Robots and Systems 2013.

Iterative Learning for Periodic Quadrocopter Maneuvers

Markus Hehn: “This video demonstrates an iterative learning algorithm that allows accurate trajectory tracking for quadrocopters executing periodic maneuvers.

The algorithm uses measurements from past executions in order to find corrections that lead to better tracking performance. In order to do this, we measure the tracking error over two laps of the maneuver. The new correction is then computed and applied. After waiting for one lap, we begin measuring again and the next learning step follows. For particularly dynamic maneuvers, we begin the learning process at lower execution speeds. This allows us to initially improve performance under safer conditions, and the algorithm provides a means to then transfer the learned corrections from the lower execution speed to higher speeds. The experience gained at lower speeds thus helps us when flying at high speeds, similar to how people learn skills such as martial arts or playing the piano. The method is also applicable to more complex tasks, shown here by the example of the quadrocopter balancing a pole while following a trajectory.”

See Markus’ research paper, presented at the IEEE/RSJ International Conference on Intelligent Robots and Systems 2013.

1. The Handbook of Robotics is a massive book. But is amazingly good at giving a broad sweep of the field.
2. Robotics, Vision and Control: Fundamental Algorithms in MATLAB by Peter Corke is an excellent introductory text with online Matlab examples.
3. Probabilistic Robotics by Sebastian Thrun et al. gives an excellent grounding in statistical and machine learning methods for robotics.
4. For people starting out in robotics I recommend the brilliant LEGO MINDSTORMS by Laurens Valk. It provides practical ways to teach and inspire the next generation of roboticists.
5. I also recommend people read some inspiring science fiction to get motivated! You can’t go past Isaac Asimov and his classic I, Robot.

Read more answers →

Machines can make use of senses we don’t possess or which are more sensitive than those we do. Their vision can extend into the infrared and ultraviolet, as well as more finely dividing the visible spectrum, and can also be more detailed and quicker (tracking faster motion) or more accurately track changes over a period of days or weeks. Their hearing can be far sharper than our own. They can be equipped with chemical sensitivity capable of distinguishing between substances we would group together under broad categories, like sweet or acrid. They can also be equipped with radar and sonar, laser ranging and scanning, accurate measures of temperature, humidity, and insolation, and their manipulators can be made to gauge and control pressure more accurately than do our own fingertips. In short, machines can have far better data available to them than would an unassisted human gardener in the same position.

Machines can also correlate information very quickly, drawing on recorded data and expert systems to make decisions, and applying heuristics to experience to refine those expert systems. A machine might reasonably be expected to identify to species every plant within the area it was tending, to know whether they were considered crops, benign, weeds, or threatened or endangered, and treat them accordingly. It might be expected to predict to an accuracy of a few days when it could harvest a particular crop, and estimate to within a few percentage points the quantity that could be expected, barring a calamity such as hail or a tornado. It might also be expected to adapt a cropping plan to market conditions, for example putting in more of some crop that hadn’t done well elsewhere and would therefore be in demand.

Machines can whisper to each other, via radio links, over distances far greater than a human shout will carry. They can coordinate their activities precisely, cooperating toward a common goal without so much as a hiccup.

Machines can, as has recently been demonstrated by DARPA’s autonomous vehicle competitions, operate in an uncontrolled environment.

The foregoing is intended as a glimpse of how it might work once development was far along. It presumes a mature technology, some of the pieces of which aren’t yet available or only just beginning to be so.

Reposted from Cultibotics.