Abstract
In this thesis I present a number of computational models and biologically interpretable agents that embody theoretical constructs of neural computation, namely the Free Energy Principle and Active Inference. The Free Energy Principle is a unifying theory of the brain that integrates concepts from statistical physics, theoretical neuroscience and machine learning; its central claim, that any self-organizing system at equilibrium with its environment must minimize its free energy, is supported by a mathematical formulation of how adaptive systems resist a natural tendency to disorder. Active Inference is a corollary of the FreeEnergy Principle that explains how action, perception and learning are unified by this objective.
Using the discrete-time Markovian formulation of Active Inference, I show that individual differences in behaviour can be explained in terms of model parameters rather than (or in addition to) traditional behavioural metrics, allowing one to make meaningful comparisons between real and simulated behaviours. I also use a model of predictive coding to simulate perceptual inference, drawing a comparison between learned visual representations in the
brain and the latent variables of deep generative models. In the final chapter, I combine these predictive coding and decision-making models to simulate both categorical perception and saccade planning in a novel visual search task. This novel contribution demonstrates that complex task stimuli can be accounted for explicitly within active inference models, enabling their use in a wider range of experimental paradigms.
Date of Award | 2 Dec 2021 |
---|---|
Original language | English |
Awarding Institution |
|
Supervisor | Nathan F Lepora (Supervisor) & Rosalyn Moran (Supervisor) |