6.891: Learning in Worlds with Objects



Most machine-learning techniques assume a representation of states of the problem in terms of vectors of discrete or real-valued attributes. We would like to apply machine learning to more complex situations that humans would describe in terms of the objects that compose them (chairs, tables, books, people). However, it is not clear how to represent such domains in order to afford effective generalization, nor what algorithms to use to learn these more complex representations.

This course will be a seminar in which we study research papers from related areas and try to develop one or more approaches to representation and learning in worlds with objects.

Syllabus

Feb 7: Varieties of learning and representation

Required: Optional:

Feb 14: Deterministic worlds with hidden state

Required: Optional:

Feb 21: Stochastic worlds with hidden state

Required: Optional:

Feb 29 (Friday!!): Deterministic relational representations

March 7: Inductive logic programming

Required: Optional:

March 14: Situated logical learning

Required:

March 21: Probabilistic relational representations

Required: Optional:

March 28: Spring Break

April 4: Probabilistic relational learning

April 11: More probabilistic relational learning

Required: Optional:

April 12 (Friday!!): Logic and the real world: Symbol grounding

April 18: No class!

April 25: More grounding, for real

May 2: Signals to symbols: learning to interpret your sensors

May 9: Signals to symbols: object representation

May 10: Project Presentations

May 16: Project presentations