Advancing models of the visual system using spiking neural networks optimized for temporal prediction

Aug 6, 2026

Upload by Finn

Spikes are treated as a fundamental computation unit in the nervous system, motivating a biologically realistic spiking neural network to study how visual information is represented by different spike patterns. The work models both retina and primary visual cortex using temporal prediction of future sensory input under metabolic-like constraints. The models reproduce V1-like spiking statistics, generate simple/complex cell tuning, capture retina-like receptive fields and phenomena such as latency coding, response omissions, and motion tuning, and better predict ganglion cell responses across species. A simulation-training acceleration method delivers a 10–50× speedup while matching benchmark performance.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages164

Reading Time6 h 53 min

AccessFree to download

Introduction

Creating an artificial visual system

Overview of the visual system

Characterizing neural tuning

The retina

Primary visual cortex

Modelling the visual system

Model architecture

Model objective: theories of visual processing

Model training: learning the neural parameters and connectivity

Model validation

Model shortcomings and new frontiers

Thesis overview

Beyond deep learning: a neuro-inspired approach to modelling primary visual cortex

Introduction

Results

Discussion

Methods

Crystal balls as eyes: retina optimised for prediction across animal species

Introduction

Results

Advancing models of the visual system using spiking neural networks optimized for temporal prediction Luke Taylor Hertford College, Oxford University Department of Physiology Anatomy & Genetics Thesis submitted for the degree of Doctor of Philosophy Hilary 2023 Abstract Spikes are thought to provide a fundamental unit of computation in the nervous system. Theretina is known to use the relative timing of spikes to encode visual input, whereas primary visual cortex (V1) exhibits sparse and irregular spiking activity – but what do these different spiking patterns represent about sensory stimuli? To address this question, I set out to model theretina and V1 using a biologically-realistic spiking neural network (SNN), exploring the idea that temporal prediction underlies the sensory transformation of natural inputs. Firstly, I trained a recurrently-connected SNN of excitatory and inhibitory units to predict the sensory future in natural movies under metabolic-like constraints. This network exhibited V1-like spike statistics, simple and complex cell-like tuning, and - advancing prior studies -key physiological and tuning differences between excitatory and inhibitory neurons. Secondly, I modiﬁed this spiking network to model the retina to explore its role in visual processing. I found the model optimized for efﬁcient prediction to capture retina-like receptive ﬁelds and-in contrast to previous studies-various retinal phenomena, such as latency coding, response omissions, and motion-tuning properties. Notably, the temporal prediction model also more accurately predicts retinal ganglion cell responses to natural images and movies across various animal species. Lastly, I developed a new method to accelerate the simulation and training of SNNs, obtaining a ∼ 10 − 50× speedup, with performance on a par with the standard training approach on supervised classiﬁcation benchmarks and for ﬁtting electrophysiological recordings of cortical neurons. The retina and V1 models lay the foundation for developing normative models of increasing biological realism and link sensory processing to spiking activity, suggesting that temporal prediction is an underlying function of visual processing. This is complemented by a new approach to drastically accelerate computational research using SNNs. ii ii Acknowledgement The last three years have undoubtedly been an incredible journey. My gratitude is due to many who have guided and supported me throughout this time. Foremost, I am grateful tomy supervisors for their guidance and the academic freedom they have granted me to explore my interests. A special thank you to Nicol and Andy for the many hours you have put into this work and the many lessons you have taught me. Your attention to detail has instilled in me a greater sense of scientiﬁc rigour. I would also like to thank Friedemann for drawing me into the intriguing world of spiking neural networks. I am thankful to the Clarendon Fund, without whose generous funding my DPhil would not have been possible; and to Sarah for helping me navigate all of the university guidelines. I am immensely grateful to my family and friends. To my parents and brother for their love and encouragement over the years. To my grandparents and granny Eva for your kind messages and support. To friends old and new, who bring much joy and vibrance to my life. Especially to Adi and Chris for our continued adventures over the years – no matter where we are. And most importantly to Melissa, for your love and companionship. As scientists say, the greatest discoveries are those that you least expect and you undoubtedly mark such an occurrence in my life. iii iii Contents 1 Introduction 1 1.1 Creating an artiﬁcial visual system .......................... 1 1.2 Overview of the visual system ............................. 4 1.2.1 Characterizing neural tuning ......................... 6 1.2.2 The retina .................................... 7 1.2

FAQs

What central idea does the thesis use to model sensory transformation?

It proposes that temporal prediction underlies the sensory transformation of natural inputs, and tests this using biologically realistic spiking neural networks.

How does the spiking V1 model behave under the trained temporal prediction objective?

The trained recurrent excitatory-inhibitory network yields V1-like spike statistics and produces simple- and complex-cell-like tuning, along with physiological and tuning differences between excitatory and inhibitory neurons.

What retina-related findings does the optimized prediction model reproduce?

The retina-optimized network captures retina-like receptive fields and reproduces phenomena such as latency coding, response omissions, and motion-tuning properties, with improved prediction of ganglion cell responses across animal species.

Aug 6, 2026

Upload by Finn