Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session A: Tuesday, August 12, 1:30 – 4:30 pm, de Brug & E‑Hall

Continuous-time Bayesian causal structure learning explains dopamine and behavioral anomalies in associative learning experiments

John Vastola¹; ¹Harvard Medical School

Presenter: John Vastola

Cause-effect learning is a core competency of sufficiently intelligent animals, and has been invoked to try to explain anomalous results (in both behavior and phasic dopamine activity) in certain associative learning experiments. But it is unclear how to mathematically formalize the problem of cause-effect learning, especially in a way that accommodates conditional independence structure, priors, and temporal structure (i.e., event order and proximity in time). We propose a novel Bayesian framework for modeling cause-effect learning which incorporates each of those aspects, yet remains relatively simple and has few free parameters. We study salient mathematical properties of our framework, including how inference is affected by topological structure in the assumed causal graph. Finally, we apply our framework to explain associative learning experiments, and find that it parsimoniously accounts for many otherwise puzzling observations. For example, our model explains the observation that cue-reward associations can be weakened by providing free reward at other times (contingency degradation), but only if the free reward is uncued. It also explains the observation that associations can be learned in fewer trials if each trial is longer. Our results suggest a new way to think about cause-effect learning, and support the idea that animals exploit nontrivial (causal) state representations even in simple associative learning settings.

Topic Area: Memory, Spatial Cognition & Skill Learning

Extended Abstract: Full Text PDF