ARC Talk: Lilian Weber, Osnabrück University: Rethinking reinforcement learning in biological agents: the role of internal states and inference in generating reward signals

Rethinking reinforcement learning in biological agents: the role of internal states and inference in generating reward signals

Humans and other animals routinely adapt their behaviour to changes in their circumstances, including the external as well as their internal environment (goals and needs). Reinforcement learning (RL) has emerged as the go-to framework to understand how goal-directed agents learn from interacting with their environment. In this talk, I will argue that the conventional formulation of RL misses a critical component of reward related processing and goal-directed action in biological agents. Standard RL models treat the reward signal as an exogenous input - something given to the learner rather than generated by it.

This assumption overlooks a crucial part of the biological intelligence: in living systems, primary rewards emerge endogenously from interoceptive inference: the ongoing process of predicting and regulating internal (bodily) states. I will review evidence from the animal and human literature on the origins and flexibility of biological reward functions to derive constraints that an updated RL model should fulfil, and present some ongoing efforts to formalise and empirically study the relevant inference mechanisms.

https://arc.mpib-berlin.mpg.de/public-with-recording/

Meeting-ID: 691 9415 1115
Kenncode: 515650