Algorithms for interactive learning

Abstract in Lay TermsAlgorithms for interactive learningSanjoy Dasgupta, UC of San-DiegoSivan Sabato, Ben-Gurion University of the NegevMachine Learning is a modern and successful approach for letting a com-puter learn from examples. However, the way machine learning is often usedtoday is limited to giving the computer examples, and for each example its cor-rect answer, for instance: photos of animals, and for each face the name of theanimal in the photo.

This approach is dierent from the way people learn: they use other infor-mation as well. For instance, if a child confuses a dog for a cat, the adult willnot only tell them \this is a dog", they will also explain why: \It barks so itcan't be a cat". Alternatively, the adult can say: \It is a dog: look how similarit is to our neighbor's dog". This type of feedback is helpful for the learningprocess of humans, and we expect it to be helpful for computer learning as well.

In this research we will study how dierent types of feedback, that are richerthan just giving the right answer for a set of examples, can help in improvingthe learning process of computers. We will create new methods that can use thistype of feedback, and also request it from a human. For instance, the computerwill be able to present a photo to a human and ask the human to tell them whatit is, but also why. For instance, the computer can ask: \Why is this a dog andnot a cat?", to which the human can provide a helpful explanation.

By developing methods that allow the computer to learn from richer typesof feedback, we expect to increase the success and availability of tools based oncomputer learning, for various tasks that today are too dicult for the com-puter to learn on its own. This will have a positive eect on the technologicalpossibilities in many elds, from applications for end-users to scientic uses. Anexample of a scientic use is, for instance, having the computer learn how toidentify microorganisms using human guidance.

1