Spring Semester

This course presents basic frameworks of learning, offering the theoretical underpinning for the development of machine learning algorithms, with an emphasis on the development of naturalistic solutions for the acquisition of symbolically-represented cognitive knowledge. It examines learning in the limit, the mistake-bounded model of online learning, active learning with queries, and the probably approximately correct model of batch learning. It then discusses learnability in the presence of missing or corrupted information. An effort is made to connect the formal properties of these models to real world situations, and examine the extent to which these properties capture or reflect some aspects of human learning. The relation of learning to the processes of perception and reasoning is also discussed, as well as the relation of learning to other natural processes, including the process of evolution. The course then examines the design of cognitive assistants as a general human-machine interaction paradigm and as a use-case for the application and interaction of perception, learning, and reasoning.

The key goals of the course are to: (i) introduce the formal theory of computational learning, and the need to develop learning solutions that can be accompanied by guarantees on their efficiency and effectiveness; (ii) present and contrast various formalizations of learning, and identify their features and the scope of their applicability; (iii) exemplify the close interaction that learning has with perception and reasoning; (iv) promote the human-machine interaction paradigm of viewing a learning machine as a cognitive assistant to a decision-maker; (v) distinguish the learning of background knowledge from the learning of user-specific policies, and identify the forms of learning that are most appropriate for each of the two cases.