University of Sussex
Browse
Chorley,_Paul.pdf (3.36 MB)

The influence of dopamine on prediction, action and learning

Download (3.36 MB)
thesis
posted on 2023-06-08, 11:50 authored by Paul Chorley
In this thesis I explore functions of the neuromodulator dopamine in the context of autonomous learning and behaviour. I first investigate dopaminergic influence within a simulated agent-based model, demonstrating how modulation of synaptic plasticity can enable reward-mediated learning that is both adaptive and self-limiting. I describe how this mechanism is driven by the dynamics of agentenvironment interaction and consequently suggest roles for both complex spontaneous neuronal activity and specific neuroanatomy in the expression of early, exploratory behaviour. I then show how the observed response of dopamine neurons in the mammalian basal ganglia may also be modelled by similar processes involving dopaminergic neuromodulation and cortical spike-pattern representation within an architecture of counteracting excitatory and inhibitory neural pathways, reflecting gross mammalian neuroanatomy. Significantly, I demonstrate how combined modulation of synaptic plasticity and neuronal excitability enables specific (timely) spike-patterns to be recognised and selectively responded to by efferent neural populations, therefore providing a novel spike-timing based implementation of the hypothetical ‘serial-compound’ representation suggested by temporal difference learning. I subsequently discuss more recent work, focused upon modelling those complex spike-patterns observed in cortex. Here, I describe neural features likely to contribute to the expression of such activity and subsequently present novel simulation software allowing for interactive exploration of these factors, in a more comprehensive neural model that implements both dynamical synapses and dopaminergic neuromodulation. I conclude by describing how the work presented ultimately suggests an integrated theory of autonomous learning, in which direct coupling of agent and environment supports a predictive coding mechanism, bootstrapped in early development by a more fundamental process of trial-and-error learning.

History

File Version

  • Published version

Pages

195.0

Department affiliated with

  • Informatics Theses

Qualification level

  • doctoral

Qualification name

  • dphil

Language

  • eng

Institution

University of Sussex

Full text available

  • Yes

Legacy Posted Date

2012-06-20

Usage metrics

    University of Sussex (Theses)

    Categories

    No categories selected

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC