Normative Hidden Variable Models of Learning and Decision Making Under Uncertainty
Mandana Ahmadi · UCL Discovery (University College London) · 2020
Understanding the mechanisms behind learning and decision making under uncertainty remains an open challenge, despite a wealth of experimental and theoretical studies. In this thesis, we focus on passive learning and perceptual decision making, as investigated in classical conditioning experiments and the random-dot-kinematogram task, respectively. We show how both problem settings can be successfully modelled in a compact and theoretically sound manner by formulating them as normative latent variable models. By being explicit about both the statistical nature of the task setting, where hidden (latent) causes have to be inferred from noisy observations, and the computational goal the animal is trying to solve, we are able to arrive at powerful models that can explain a greater variety of data with fewer assumptions and free parameters than previous approaches, and we offer predictions about how behaviour is expected to change if task parameters or neural activity are altered. In the first half of this thesis we focus on decision making, and introduce a neurally plausible model that can link the task setting to behaviour. We then extend this work to propose a systems level model that includes motor control, such that previously puzzling data from neural recordings can be reconciled. In the second half of the thesis we focus on understanding passive learning in the context of classical conditioning. We show that our model can resolve long-standing disputes and differences between existing models of classical conditioning, by showing that they can be understood as special cases of our more general formulation.