GATSBY COMPUTATIONAL NEUROSCIENCE UNIT

Rational and irrational models in classical and instrumental conditioning

Recent work interprets animal conditioning behavior in terms of principled statistical reasoning. This view sheds new light on many issues previously studied from a more phenomenological perspective, including attentional and 'associability' effects, 'configural' learning effects, and phenomena of 'generalized drive' and response vigor. However, the computations envisioned by such ideal-observer models are also typically intractable in realistic circumstances, and organisms must therefore employ computational approximations to accomplish them. Thus, while some features of previous, descriptive models may be explained from a fully normative perspective, other aspects may reflect computationally expedient shortcuts or heuristics. The tradeoff between different shortcuts may itself be subject to a rational analysis, as we demonstrate with an example from control in sequential decision tasks.