Monthly Archives: August 2017

Answer plain language business queries in minutes instead of months

Companies often employ number-crunching data scientists to gather insights such as which customers want certain services or where to open new stores and stock products. Analyzing the data to answer one or two of those queries, however, can take weeks or even months.

Now MIT spinout Endor has developed a predictive-analytics platform that lets anyone, tech-savvy or not, upload raw data and input any business question into an interface — similar to using an online search engine — and receive accurate answers in just 15 minutes.

The platform is based on the science of “social physics,” co-developed at the MIT Media Lab by Endor co-founders Alex “Sandy” Pentland, the Toshiba Professor of Media Arts and Sciences, and Yaniv Altshuler, a former MIT postdoc. Social physics uses mathematic models and machine learning to understand and predict crowd behaviors.

Users of the new platform upload data about customers or other individuals, such as records of mobile phone calls, credit card purchases, or web activity. They use Endor’s “query-builder” wizard to ask questions, such as “Where should we open our next store?” or “Who is likely to try product X?” Using the questions, the platform identifies patterns of previous behavior among the data and uses social physics models to predict future behavior. The platform can also analyze fully encrypted data-streams, allowing customers such as banks or credit card operators to maintain data privacy.

“It’s just like Google. You don’t have to spend time thinking, ‘Am I going to spend time asking Google this question?’ You just Google it,” Altshuler says. “It’s as simple as that.”

Financially backed by Innovation Endeavors, the private venture capital firm of Eric Schmidt, executive chairman of Google parent company Alphabet, Inc., the startup has found big-name customers, such as Coca-Cola, Mastercard, and Walmart, among other major retail and banking firms.

Recently, Endor analyzed Twitter data for a defense agency to detect potential terrorists. Endor was given 15 million data points containing examples of 50 Twitter accounts of identified ISIS activists, based on identifiers in the metadata. From that, they asked the startup to detect 74 with identifiers extremely well hidden in the metadata. Someone at Endor completed the task on a laptop in 24 minutes, detecting 80 “lookalike” ISIS accounts, 45 of which were from the pool of 74 well-hidden accounts named by the agency. The false positive rate was also extremely low (35 accounts), meaning that human analysts could afford to have experts investigating the accounts.

Finds patterns in materials recipes

Last month, three MIT materials scientists and their colleagues published a paper describing a new artificial-intelligence system that can pore through scientific papers and extract “recipes” for producing particular types of materials.

That work was envisioned as the first step toward a system that can originate recipes for materials that have been described only theoretically. Now, in a paper in the journal npj Computational Materials, the same three materials scientists, with a colleague in MIT’s Department of Electrical Engineering and Computer Science (EECS), take a further step in that direction, with a new artificial-intelligence system that can recognize higher-level patterns that are consistent across recipes.

For instance, the new system was able to identify correlations between “precursor” chemicals used in materials recipes and the crystal structures of the resulting products. The same correlations, it turned out, had been documented in the literature.

The system also relies on statistical methods that provide a natural mechanism for generating original recipes. In the paper, the researchers use this mechanism to suggest alternative recipes for known materials, and the suggestions accord well with real recipes.

The first author on the new paper is Edward Kim, a graduate student in materials science and engineering. The senior author is his advisor, Elsa Olivetti, the Atlantic Richfield Assistant Professor of Energy Studies in the Department of Materials Science and Engineering (DMSE). They’re joined by Kevin Huang, a postdoc in DMSE, and by Stefanie Jegelka, the X-Window Consortium Career Development Assistant Professor in EECS.

Sparse and scarce

Like many of the best-performing artificial-intelligence systems of the past 10 years, the MIT researchers’ new system is a so-called neural network, which learns to perform computational tasks by analyzing huge sets of training data. Traditionally, attempts to use neural networks to generate materials recipes have run up against two problems, which the researchers describe as sparsity and scarcity.

Any recipe for a material can be represented as a vector, which is essentially a long string of numbers. Each number represents a feature of the recipe, such as the concentration of a particular chemical, the solvent in which it’s dissolved, or the temperature at which a reaction takes place.

Since any given recipe will use only a few of the many chemicals and solvents described in the literature, most of those numbers will be zero. That’s what the researchers mean by “sparse.”

Similarly, to learn how modifying reaction parameters — such as chemical concentrations and temperatures — can affect final products, a system would ideally be trained on a huge number of examples in which those parameters are varied. But for some materials — particularly newer ones — the literature may contain only a few recipes. That’s scarcity.

“People think that with machine learning, you need a lot of data, and if it’s sparse, you need more data,” Kim says. “When you’re trying to focus on a very specific system, where you’re forced to use high-dimensional data but you don’t have a lot of it, can you still use these neural machine-learning techniques?”

Neural networks are typically arranged into layers, each consisting of thousands of simple processing units, or nodes. Each node is connected to several nodes in the layers above and below. Data is fed into the bottom layer, which manipulates it and passes it to the next layer, which manipulates it and passes it to the next, and so on. During training, the connections between nodes are constantly readjusted until the output of the final layer consistently approximates the result of some computation.

Artificial intelligence systems that process language

Neural networks, which learn to perform computational tasks by analyzing huge sets of training data, have been responsible for the most impressive recent advances in artificial intelligence, including speech-recognition and automatic-translation systems.

During training, however, a neural net continually adjusts its internal settings in ways that even its creators can’t interpret. Much recent work in computer science has focused on clever techniques for determining just how neural nets do what they do.

In several recent papers, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Qatar Computing Research Institute have used a recently developed interpretive technique, which had been applied in other areas, to analyze neural networks trained to do machine translation and speech recognition.

They find empirical support for some common intuitions about how the networks probably work. For example, the systems seem to concentrate on lower-level tasks, such as sound recognition or part-of-speech recognition, before moving on to higher-level tasks, such as transcription or semantic interpretation.

But the researchers also find a surprising omission in the type of data the translation network considers, and they show that correcting that omission improves the network’s performance. The improvement is modest, but it points toward the possibility that analysis of neural networks could help improve the accuracy of artificial intelligence systems.

“In machine translation, historically, there was sort of a pyramid with different layers,” says Jim Glass, a CSAIL senior research scientist who worked on the project with Yonatan Belinkov, an MIT graduate student in electrical engineering and computer science. “At the lowest level there was the word, the surface forms, and the top of the pyramid was some kind of interlingual representation, and you’d have different layers where you were doing syntax, semantics. This was a very abstract notion, but the idea was the higher up you went in the pyramid, the easier it would be to translate to a new language, and then you’d go down again. So part of what Yonatan is doing is trying to figure out what aspects of this notion are being encoded in the network.”

Sensors could be sensitive enough for self-driving cars

For the past 10 years, the Camera Culture group at MIT’s Media Lab has been developing innovative imaging systems — from a camera that can see around corners to one that can read text in closed books — by using “time of flight,” an approach that gauges distance by measuring the time it takes light projected into a scene to bounce back to a sensor.

In a new paper appearing in IEEE Access, members of the Camera Culture group present a new approach to time-of-flight imaging that increases its depth resolution 1,000-fold. That’s the type of resolution that could make self-driving cars practical.

The new approach could also enable accurate distance measurements through fog, which has proven to be a major obstacle to the development of self-driving cars.

At a range of 2 meters, existing time-of-flight systems have a depth resolution of about a centimeter. That’s good enough for the assisted-parking and collision-detection systems on today’s cars.

But as Achuta Kadambi, a joint PhD student in electrical engineering and computer science and media arts and sciences and first author on the paper, explains, “As you increase the range, your resolution goes down exponentially. Let’s say you have a long-range scenario, and you want your car to detect an object further away so it can make a fast update decision. You may have started at 1 centimeter, but now you’re back down to [a resolution of] a foot or even 5 feet. And if you make a mistake, it could lead to loss of life.”

At distances of 2 meters, the MIT researchers’ system, by contrast, has a depth resolution of 3 micrometers. Kadambi also conducted tests in which he sent a light signal through 500 meters of optical fiber with regularly spaced filters along its length, to simulate the power falloff incurred over longer distances, before feeding it to his system. Those tests suggest that at a range of 500 meters, the MIT system should still achieve a depth resolution of only a centimeter.

Kadambi is joined on the paper by his thesis advisor, Ramesh Raskar, an associate professor of media arts and sciences and head of the Camera Culture group.

Slow uptake

With time-of-flight imaging, a short burst of light is fired into a scene, and a camera measures the time it takes to return, which indicates the distance of the object that reflected it. The longer the light burst, the more ambiguous the measurement of how far it’s traveled. So light-burst length is one of the factors that determines system resolution.

The other factor, however, is detection rate. Modulators, which turn a light beam off and on, can switch a billion times a second, but today’s detectors can make only about 100 million measurements a second. Detection rate is what limits existing time-of-flight systems to centimeter-scale resolution.