An analysis and extension of predictive coding for learning in the brain

Thesis

Understanding how the brain learns is a central objective of neuroscience, with important consequences for neurological disease, rehabilitation and education. Predictive coding has emerged as a compelling candidate for a unifying theory of brain function. It proposes that the brain learns by minimising the discrepancy between incoming sensory signals and its own predictions. In doing so, it offers a biologically plausible account of learning that respects three key constraints of neural systems: hierarchical organisation, local information processing and strong learning performance. Despite these strengths, predictive coding in its standard form does not fully capture the breadth of abilities exhibited by the brain within a single network. This thesis extends predictive coding along three complementary directions. First, we show how predictive coding can represent uncertainty, a capacity essential for robust perception and decision-making. We introduce Monte Carlo predictive coding, which augments predictive coding’s activity dynamics to estimate full posterior distributions while preserving local computations. Second, we demonstrate how predictive coding can support both supervised and unsupervised learning within a biologically plausible architecture that reflects computational motifs observed in visual perception. In this framework, each neural layer predicts the activity of both the layer above and the layer below in the hierarchy, enabling flexible learning across tasks. Third, we develop an analytical account explaining why predictive coding can learn with reduced interference, mirroring the brain’s ability to acquire new knowledge with minimal catastrophic forgetting. We identify the regimes in which these advantages are most pronounced and propose modifications that guarantee, or closely approximate, interference-free learning. Together, these contributions expand the functional scope of predictive coding and clarify the conditions under which its learning dynamics are both powerful and robust.

Authors: Oliviers, G

• DOI: 10.5287/ora-erogjjzd6
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English