Human-Level Learning

Computer and cognitive scientists at New York University has been able to demonstrate the ability of AI (artificial intelligence) to quickly grasp a new concept after observing a single example, and it can generate new ideas from this experience. The program is large part funded by the U.S. military and could greatly impact the military’s ability to analyze data.

Published today in the journal Science, the paper “Human-level Concept Learning Through Probabilistic Program Induction” asserts that, remarkably, the AI is able to learn at the same speed as humans, at least in relation to certain concepts. Ultimately, the paper presents a model called the Bayesian Program Learning (BPL) framework. It is thanks to this framework that this program is capable of classifying objects and generating concepts on those objects using only a single instance.

In their test, the research team showed BPL, along with several other people, 20 handwritten letters from 10 different alphabets. The subjects were then asked to match the given letter with the same character written by someone else. The results have BPL scoring 97%, which is about as high as humans score on this test.

Fig. 1. People can learn rich concepts from limited data. A single example of a new concept (red boxes) can be enough information to support the (i) classification of new examples, (ii) generation of new examples, (iii) parsing an object into parts and relations (parts segmented by color), and (iv) generation of new concepts from related concepts.
Evolution of Machine Learning

"The core of the program is probabilistic—every time you run it it produces a different [interpretation]," says co-author Brenden M. Lake in the press release.

"There's still a big gap between what people are capable of and what we can do with machine learning today, but our program has captured a range of new learning capabilities," which are human-like in "their ability to generalize from a single example in much more powerful ways," says Lake.

Share This Article