Visual cortex is optimised for short timescale prediction using spikes


A key question in systems neuroscience is to understand what principles underly the sensory processing throughout the brain. Why are certain neurons in V1 selectively tuned to orientated bars? Or why is V1 spike activity mostly sparse and irregular? During my PhD I have been investigating the central hypothesis that the brain is optimised to encode features of the recent sensory past that are predictive of the immediate sensory future. Recent modelling work using artificial neural networks (ANNs) for prediction on natural stimuli have accounted for various biological phenomena in the visual system. However, these models omit key physiological constraints, such as the spiking nature of neurons, excitatory and inhibitory neural subpopulations and axonal conductance latencies. Including these constraints, and training a recurrently connected spiking neural network (SNN) on natural video stimuli – with the objective of predicting video frames 40ms into the future – successfully reproduces a number of biological phenomena found in V1, such as different receptive fields, spike statistics and neural tuning properties. These findings support the hypothesis that the visual circuit is optimised to enable short timescale prediction.


Luke Taylor I am a neuroscience PhD student at Oxford University, where my research has dichotomously been focused on spiking neural networks (SNNs): 1. Using these networks to study the brain and 2. Developing methods to accelerate the training and improve the accuracy of these networks. Before pursuing my PhD, I completed my MSc in Neuroscience at Oxford, and my honours and undergraduate in Applied Mathematics at the University of Cape Town.


23 November 2022