10-812 Architectures for Never-Ending Learning
12 units, Carnegie Mellon University
This upper level graduate seminar course will examine the question of how to design software architectures for neverending learning agents. We will build on early attempts to do this, studying systems including the Never Ending Language Learner NELL, and Never Ending Image Learner NEIL. But our goal is not to simply understand existing systems it is to design the next generation of never ending learners. Key questions we will consider include What architectural components are needed by a sensoreffector agent, to support never ending learning How can such an agent avoid performance plateaus, in which it reaches a level of competence beyond which it can no longer improve What components and mechanisms beyond generic deep networks need to be incorporated into the architecture What is the role of selfreflection What types of representation learning must be incorporated What are the roles for supervised training, conversations, and unsupervised experimentation in never ending learning and What sequence of learning tasks form an appropriate curriculum for the learning agent to learn to learn The format of the course will include group discussions of research papers read in advance, presentations on research topics by students, guest lectures, and a course project in which student teams design and experiment with their own never ending learner. Prerequisite ML department graduate course in machine learning 10715, 10701 or 10601, plus permission of the instructor.
Prerequisites: 10-701 (min C) or 10-715 (min C) or 10-601 (min C)
Courseweave needs JavaScript to show sections, workloads and planning tools.