Vertes, Eszter;
(2020)
Probabilistic learning and computation in brains and machines.
Doctoral thesis (Ph.D), UCL (University College London).
Preview |
Text
Vertes_thesis.pdf Download (6MB) | Preview |
Abstract
Humans and animals are able to solve a wide variety of perceptual, decision making and motor tasks with great exibility. Moreover, behavioural evidence shows that this exibility extends to situations where accuracy requires the correct treatment of uncertainty induced by noise and ambiguity in the available sensory information as well as noise internal to the brain. It has been suggested that this adequate handling of uncertainty is based on a learned internal model, e.g. in the case of perception, a generative model of sensory observations. Learning latent variable models and performing inference in them is a key challenge for both biological and arti cial learning systems. Here, we introduce a new approach to learning in hierarchical latent variable models called the Distributed Distributional Code Helmholtz Machine (DDC-HM), which emphasises exibility and accuracy in the inferential process. The approximate posterior over unobserved variables is represented implicitly as a set of expectations, corresponding to mean parameters of an exponential family distribution. To train the generative and recognition models we develop an extended wake-sleep algorithm inspired by the original Helmholtz Machine. As a result, the DDC-HM is able to learn hierarchical latent models without having to propagate gradients across di erent stochastic layers|making our approach biologically appealing. In the second part of the thesis, we review existing proposals for neural representations of uncertainty with a focus on representational and computational exibility as well as experimental support. Finally, we consider inference and learning in dynamical environment models using Distributed Distributional Codes to represent both the stochastic latent transition model and the inferred posterior distributions. We show that this model makes it possible to generalise successor representations to biologically more realistic, partially observed settings.
| Type: | Thesis (Doctoral) |
|---|---|
| Qualification: | Ph.D |
| Title: | Probabilistic learning and computation in brains and machines |
| Event: | UCL |
| Open access status: | An open access version is available from UCL Discovery |
| Language: | English |
| Additional information: | Copyright © The Author 2020. Original content in this thesis is licensed under the terms of the Creative Commons Attribution 4.0 International (CC BY 4.0) Licence (https://creativecommons.org/licenses/by/4.0/). Any third-party copyright material present remains the property of its respective owner(s) and is licensed under its existing terms. Access may initially be restricted at the author’s request. |
| UCL classification: | UCL UCL > Provost and Vice Provost Offices UCL > Provost and Vice Provost Offices > School of Life and Medical Sciences UCL > Provost and Vice Provost Offices > School of Life and Medical Sciences > Faculty of Life Sciences |
| URI: | https://discovery-pp.ucl.ac.uk/id/eprint/10103090 |
Archive Staff Only
 |
View Item |
