Exploring normative models of the visual and auditory systems

Thesis

The brain’s sensory systems transform the incoming stimuli which impinge on the periphery into useful representations that can guide an organism’s behaviour. Indeed, understanding the complex neural computations which underpin these transformations across sensory pathways remains an enduring goal of systems neuroscience. Normative modelling provides one approach to tackling this complexity, by describing neural systems from an optimization perspective. In this way, a normative approach can answer to what extentthe diverse structural and functional properties of the sensory brain might emerge by optimizing for a few more fundamental principles of neural function. This normative approach forms the core of this thesis, which I explore across the visual and auditory systems.

To that end, in the first two results chapters, I investigate how optimizing recurrent artificial neural networks for predictive information – termed temporal prediction – can capture both the structural and functional properties of the mammalian visual system. Specifically, in chapter two, I describe how a shallow recurrent model trained for temporal prediction can recapitulate the functional connectivity motifs of mouse primary visual cortex (V1). In chapter three, I extend this V1 model into a hierarchical recurrent model of the dorsal visual pathway. There, I demonstrate how feedback connectivity in the network captures many of the known functional properties of higher-order feedback to V1. Finally, in chapter four, I move from the visual system to the auditory system where I take a broader view across the wider landscape of normative models. In this chapter, I investigate which properties determine how well normative models can predict neural activity and how this relates to the models’ learned representations. In particular, I show that networks which learn more general representations are better able to model auditory neural activity.

Overall, this thesis demonstrates the utility of normative networksas models of the brain and shows how the complexity sensory systems might emerge by optimizing for much simpler principles such as temporal prediction.

Authors: Klavinskis-Whiting, S

• DOI: 10.5287/ora-aqypeapn0
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: auditory; visual; temporal prediction; neuroscience
• Subjects: Computational Neuroscience