What do Deep Neural Networks Tell us about Biological Vision?
Professor of Cognitive Neuroscience
Dean, Institute of Life and Human Science
University of Liverpool
Venue| Senior Common Room, Level 2, Priory Road Complex
Date | Thursday 24 October 2019
Time | 13:00
Recent years have seen a massive surge of interest in the potential application of deep learning and deep neural networks (DNNs) to a broad range of issues in biological vision. Indeed, DNNs have been described by some as a new framework for vision research. These claims are, in part, based on work showing human-level performance by DNNs in tasks such as image classification, and are supported by advances in techniques for comparing ‘representational’ structures computed by DNNs with behavioural and neuroimaging data. However, the suitability of DNNs as a theoretical framework for understanding biological vision remains unclear. In this talk, I will explore this question – and present a critical analysis of DNNs as models of human vision in the context of pattern classification and 3D object recognition. I ask several key questions: ‘How should theoretically relevant and irrelevant parameters of DNNs be distinguished?’, ‘Can DNN states and outputs be rigorously and meaningfully compared to human performance?’ ‘What is the range of empirical phenomena that must be considered to evaluate DNN architecture and processing parameters?’ I will illustrate these points with reference to recent work examining patterns of errors made by DNNs during image classification, visual illusions and the discrimination of possible and impossible objects. I argue that feed-forward, data-driven, deep learning approaches do not provide a sufficiently rich framework for elucidating the functional architecture of human vision in these domains.