Thinking is Another Form of Control

Dual process theory divides cognitive processing into a fast, intuitive System 1 and a slow, deliberative System 2. In reinforcement learning (RL), model-free learning, in which the agent takes actions with a reactive policy, is reminiscent of System 1, whereas model-based decision-time planning is reminiscent of System 2. In this talk, I will present the view that deliberative, System 2 behaviors ("thinking") can be considered a form of mental action that an agent performs before taking an action that influences its external environment. Under this view, model-free RL alone could be sufficient to produce deliberation if these mental actions ultimately led to higher value actions being selected. I'll formalize the notion of a controllable “thought" state, then prove conditions under which "thinking" emerges as a strategy for reward maximization, and discuss how large language models serve as a proof-of-concept for thinking as mental action. Finally, I'll conclude by discussing on-going work developing agents that learn when to think and act.

(This talk is part of the weekly Machine Learning Lunch Meetings (MLLM), held every Tuesday from 12:15 to 1:15 p.m.  Professors from Computer Sciences, Statistics, ECE, the iSchool, and other departments will discuss their latest research in machine learning, covering both theory and applications. This is a great opportunity to network with faculty and fellow researchers, learn about cutting-edge research at our university, and foster new collaborations. For the talk schedule, please visit https://sites.google.com/view/wiscmllm/home. To receive future weekly talk announcements, please subscribe to our UW Google Group at https://groups.google.com/u/1/a/g-groups.wisc.edu/g/mllm.)