Giving machine learning systems

Machine learning, which is the basis for most commercial artificial-intelligence systems, is intrinsically probabilistic. An object-recognition algorithm asked to classify a particular image, for instance, might conclude that it has a 60 percent chance of depicting a dog, but a 30 percent chance of depicting a cat.

At the Annual Conference on Neural Information Processing Systems in December, MIT researchers will present a new way of doing machine learning that enables semantically related concepts to reinforce each other. So, for instance, an object-recognition algorithm would learn to weigh the co-occurrence of the classifications “dog” and “Chihuahua” more heavily than it would the co-occurrence of “dog” and “cat.”

In experiments, the researchers found that a machine-learning algorithm that used their training strategy did a better job of predicting the tags that human users applied to images on the Flickr website than it did when it used a conventional training strategy.

“When you have a lot of possible categories, the conventional way of dealing with it is that, when you want to learn a model for each one of those categories, you use only data associated with that category,” says Chiyuan Zhang, an MIT graduate student in electrical engineering and computer science and one of the new paper’s lead authors. “It’s treating all other categories equally unfavorably. Because there are actually semantic similarities between those categories, we develop a way of making use of that semantic similarity to sort of borrow data from close categories to train the model.”

Zhang is joined on the paper by his thesis advisor, Tomaso Poggio, the Eugene McDermott Professor in the Brain Sciences and Human Behavior, and by his fellow first author Charlie Frogner, also a graduate student in Poggio’s group. Hossein Mobahi, a postdoc in the Computer Science and Artificial Intelligence Laboratory, and Mauricio Araya-Polo, a researcher with Shell Oil, round out the paper’s co-authors.

Close counts

To quantify the notion of semantic similarity, the researchers wrote an algorithm that combed through Flickr images identifying tags that tended to co-occur — for instance, “sunshine,” “water,” and “reflection.” The semantic similarity of two words was a function of how frequently they co-occurred.

Ordinarily, a machine-learning algorithm being trained to predict Flickr tags would try to identify visual features that consistently corresponded to particular tags. During training, it would be credited with every tag it got right but penalized for failed predictions.

The MIT researchers’ system essentially gives the algorithm partial credit for incorrect tags that are semantically related to the correct tags. Say, for instance, that a waterscape was tagged, among other things, “water,” “boat,” and “sunshine.” With conventional machine learning, a system that tagged that image “water,” “boat,” “summer” would get no more credit than one that tagged it “water,” “boat,” “rhinoceros.” With the researchers’ system, it would, and the credit would be a function of the likelihood that the tags “summer” and “sunshine” co-occur in the Flickr database.