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.

Photo credit: Robert Shields

Will your next doctor be an app? A cost-cutting NHS wants more patients to act as “self-carers,” with some technologized assistance. A series of flowcharts and phone trees might tell parents whose children have chicken pox how best to care for them—no visits to surgeries required. Or a mole-checking app might tell a worrywart when a given skin discoloration looks harmless, and when to go to a dermatologist, by comparing it to thousands of images in a database.

Cost-cutters in the legal field also promise an algorithmically cheapened future. Tax software simplifies the process of filing by walking the filer through a series of questions. Documents that might have taken human attorneys months to read, can be scanned for keywords in a matter of seconds. Predictive policing promises to deploy force with surgical precision.

All these initiatives have some promise, and may make health care and legal advice more accessible. But they are also prone to errors, biases, and predictable malfunctions. Last year, the US Federal Trade Commission settled lawsuits against firms who claimed their software could aid in the detection of skin cancer, by evaluating photographs of the user’s moles. The FTC argued that there was not sufficient evidence to support such claims. The companies are now prohibited from making any “health or disease claims” about the impact of the apps on the health of users unless they provide “reliable scientific evidence” grounded in clinical tests. If algorithms designed merely to inform patients aren’t ready for prime time, why presume diagnostic robots are imminent?

Legal automation has also faced some serious critiques lately. The University of North Carolina legal scholar Dana Remus has questioned the value and legitimacy of the “predictive coding” now deployed in many discovery proceedings. She and co-author Frank S. Levy (of MIT) raise serious questions about more advanced applications of legal automation as well. The future cannot be completely anticipated in contracts, nor can difficult judgment calls be perfectly encoded into the oft-reductionist formulae of data processing. Errant divorce software may have caused thousands of errors in the UK lately, just as US software systems have disrupted or derailed proper dispositions of benefits applications.

Moreover, several types of opacity impede public understanding of algorithmic ranking and rating processes in even more familiar contexts, like credit scoring or search rankings. Consumers do not understand all the implications of the US credit scoring process, and things are about to get worse as “alternative” or “fringe” data moves into the lending mix for some startups. If the consequences of being late on a bill are not readily apparent to consumers, how can they hope to grasp new scoring systems that draw on their social media postings, location data, and hundreds of other data points? At the level of companies, many firms do not feel that Google, Facebook, and Amazon are playing a fair game in their algorithmic rankings of websites, ads, and products. These concerns, too, are stymied by widespread secrecy of both algorithms and the data fed into them.

In response, legal scholars have focused on remediable legal secrecy (curbing trade secrets and improving monitoring by watchdogs) and complexity (forbidding certain contractual arrangements when they become so complicated that regulators or citizens cannot understand their impact). I have recommended certain forms of transparency for software—for example, permitting experts to inspect code at suspect firms, and communications between managers and technical staff. The recent Volkswagen scandal served as yet another confirmation of the need for regulators to understand code.

But there is a larger lesson in these failures of algorithmic ordering. Rather than trying to replace the professions with robots and software, we should instead ask how professional expertise can better guide the implementation of algorithmic decision-making procedures. Ideally, doctors using software in medical settings should be able to inspect the inputs (data) that go into them, restrict the contexts in which they are used, and demand outputs that avoid disparate impacts. The same goes for attorneys, and other professionals now deploying algorithmic arrangements of information. We will be looking at “The Promise and Limits of Algorithmic Accountability in the Professions” at Yale Law School this Spring, and welcome further interventions to clarify the complementarity between professional and computational expertise.

This post was originally published on the website of Nesta.

Stills from the Weather Research and Forecasting Model. Image: Wikipedia/Almoz

Jennifer Chu | MIT News Office 

To work with computational models is to work in a world of unknowns: Models that simulate complex physical processes — from Earth’s changing climate to the performance of hypersonic combustion engines — are staggeringly complex, sometimes incorporating hundreds of parameters, each of which describes a piece of the larger process.

Parameters are often question marks within their models, their contributions to the whole largely unknown. To estimate the value of each unknown parameter requires plugging in hundreds, if not thousands, of values, and running the model each time to narrow in on an accurate value — a computation that can take days, and sometimes weeks.

Now MIT researchers have developed a new algorithm that vastly reduces the computation of virtually any computational model. The algorithm may be thought of as a shrinking bull’s-eye that, over several runs of a model, and in combination with some relevant data points, incrementally narrows in on its target: a probability distribution of values for each unknown parameter.

With this method, the researchers were able to arrive at the same answer as a classic computational approaches, but 200 times faster.

Youssef Marzouk, an associate professor of aeronautics and astronautics, says the algorithm is versatile enough to apply to a wide range of computationally intensive problems.

“We’re somewhat flexible about the particular application,” Marzouk says. “These models exist in a vast array of fields, from engineering and geophysics to subsurface modeling, very often with unknown parameters. We want to treat the model as a black box and say, ‘Can we accelerate this process in some way?’ That’s what our algorithm does.”

Marzouk and his colleagues — recent PhD graduate Patrick Conrad, Natesh Pillai from Harvard University, and Aaron Smith from the University of Ottawa — have published their findings this week in the Journal of the American Statistical Association.

Modeling “Monopoly”

In working with complicated models involving multiple unknown parameters, computer scientists typically employ a technique called Markov chain Monte Carlo (MCMC) analysis — a statistical sampling method that is often explained in the context of the board game “Monopoly.”

To plan out a monopoly, you want to know which properties players land on most often — essentially, an unknown parameter. Each space on the board has a probability of being landed on, determined by the rules of the game, the positions of each player, and the roll of two dice. To determine the probability distribution on the board — the range of chances each space has of being landed on — you could roll the die hundreds of times.

If you roll the die enough times, you can get a pretty good idea of where players will most likely land. This, essentially, is how an MCMC analysis works: by running a model over and over, with different inputs, to determine a probability distribution for one unknown parameter. For more complicated models involving multiple unknowns, the same method could take days to weeks to compute an answer.

Shrinking bull’s-eye

With their new algorithm, Marzouk and his colleagues aim to significantly speed up the conventional sampling process.

“What our algorithm does is short-circuits this model and puts in an approximate model,” Marzouk explains. “It may be orders of magnitude cheaper to evaluate.”

The algorithm can be applied to any complex model to quickly determine the probability distribution, or the most likely values, for an unknown parameter. Like the MCMC analysis, the algorithm runs a given model with various inputs — though sparingly, as this process can be quite time-consuming. To speed the process up, the algorithm also uses relevant data to help narrow in on approximate values for unknown parameters.

In the context of “Monopoly,” imagine that the board is essentially a three-dimensional terrain, with each space represented as a peak or valley. The higher a space’s peak, the higher the probability that space is a popular landing spot. To figure out the exact contours of the board — the probability distribution — the algorithm rolls the die at each turn and alternates between using the computationally expensive model and the approximation. With each roll of the die, the algorithm refers back to the relevant data and any previous evaluations of the model that have been collected.

At the beginning of the analysis, the algorithm essentially draws large, vague bull’s-eyes over the board’s entire terrain. After successive runs with either the model or the data, the algorithm’s bull’s-eyes progressively shrink, zeroing in on the peaks in the terrain — the spaces, or values, that are most likely to represent the unknown parameter.

“Outside the normal”

The group tested the algorithm on two relatively complex models, each with a handful of unknown parameters. On average, the algorithm arrived at the same answer as each model, but 200 times faster.

“What this means in the long run is, things that you thought were not tractable can now become doable,” Marzouk says. “For an intractable problem, if you had two months and a huge computer, you could get some answer, but you would not necessarily know how accurate it was. Now for the first time, we can say that if you run our algorithm, you can guarantee that you’ll find the right answer, and you might be able to do it in a day. Previously that guarantee was absent.”

Marzouk and his colleagues have applied the algorithm to a complex model for simulating movement of sea ice in Antarctica, involving 24 unknown parameters, and found that the algorithm is 60 times faster arriving at an estimate than current methods. He plans to test the algorithm next on models of combustion systems for supersonic jets.

“This is a super-expensive model for a very futuristic technology,” Marzouk says. “There might be hundreds of unknown parameters, because you’re operating outside the normal regime. That’s exciting to us.”

This research was supported, in part, by the Department of Energy.

Reprinted with permission of MIT News.

Manual programming of robots only gets you so far. And, as you can see in the video, for quadrocopters that’s not very far at all (see the “Without Learning” part starting at 1:30):

On its first try to navigate the obstacle course, the flying robot attempts to navigate based on a pre-computed flight path. The path is derived using a basic mathematical (“first principles”) model. But quadrocopters have complex aerodynamics – the force produced by the propellers changes depending on the vehicle’s velocity and orientation, and thus the actual amount of force produced is quite different from what the simple math describes.

What’s worse, these flying vehicles use soft propellers for safety, which bend differently depending on how much thrust is applied and wear rapidly with use (and even more rapidly when crashing).

Even with continuous feedback on the robot’s position from the motion capture system, manually programming the robots with a control sequence that takes all these imperfections into account is impractical.

My colleague Angela Schoellig and the Flying Machine Arena team here at ETH Zurich have now developed and implemented algorithms that allow their flying robots to race through an obstacle parcours – and learn to improve their performance.

Here is how Angela described the process to me:

The learning algorithm is applied to a quadrocopter that is guided by an underlying trajectory-following controller. The result of the learning is an adapted input trajectory in terms of desired positions. The algorithm has been equipped with several unique features that are particularly important when pushing vehicles to the limits of their dynamic capabilities and when applying the learning algorithm to increasingly complex systems:

1. We designed an input update rule that explicitly takes actuation and sensor limits into account by solving a constrained convex problem.
2. We developed an identification routine that extracts the model data required by the learning algorithm from a numerical simulation of the vehicle dynamics. That is, the algorithm is applicable to systems for which an analytical model is difficult (or impossible) to derive.
3. We combined model data and experimental data, traditional filtering methods and state-of-the- art optimization techniques to achieve an effective and computationally efficient learning strategy that achieves convergence in less than ten iterations.

The result is a robot that learns and improves each time it tries to perform a task.

In this example the robot races through a pylon parcours, calling to mind air races, such as the Red Bull Air Racing Championships or the Reno Air Races – except that there are no human pilots that spent their life learning to fly – here it’s robots doing the learning. And they are efficient, taking less than 10 training sessions to find the optimal steering commands!

Moreover, the learning algorithms are not specific to slalom racing, they can be used to learn other tasks. As Angela points out:

Our goal is to enable autonomous systems – such as the quadrocopter in the video – to ‘learn’ the way humans do: through practice.

The videos below show how learning algorithms can be used for other robotic tasks:

Full disclosure: Angela and the Flying Machine Arena team work in the same lab as I. Also, I’m working on RoboEarth, trying to allow robot learning on a much larger scale.