Representation in Neural Networks

Abstract

Artificial neural networks (ANNs) are computational systems that were inspired by biological neural networks in the brain. ANNs are trained to transform input into task appropriate output using learning algorithms rather than having all relevant aspects of the task explicitly encoded with symbolic rules. Despite the increasingly impressive performance and wide spread usage of ANNs in artificial intelligence (Krizhevsky et al.,2012., LeCun, et. al., 2015., Senjnowski, 2018., Floridi & Chiriatti, 2020), their operation remains somewhat mysterious. There is no widely accepted and comprehensive explanation of how these systems represent and process information (Bornstein, 2016., Habbis et. al., 2017, Schwartz-Ziv & Tishby, 2017). Approaches to explaining the operation of relatively simple neural network models have been discussed by philosophers since the inception of connectionist cognitive science. However, these discussions often relied on analysing the behaviour of a very small number of actual ANNs and there are important issues that still haven’t been resolved. I address this by using empirical data from my own unique analysis of a broad range of novel ANNs to evaluate some key philosophical approaches to understanding and comparing neural network models. I focus on structural-resemblance approaches and there has been a shift towards using structural approaches in cognitive neuroscience (Williams & Colling 2018., Boone & Piccinini, 2015., Kriegeskorte et al., 2008., Raizada and Conolly, 2012). Structural-resemblance explanations of representation rely on the idea that structural relations between representations might systematically correspond to the structural organisation of relevant aspects of the represented domain. My empirical work begins by extending Laakso and Cottrell’s (2000) method for assessing representation similarity in ANNs to explicitly compare the relevant structural relations between representations across distinct ANNs with diversely configured parameters. I apply this method to evaluate structural approaches to understanding representation in neural networks described by Churchland (1998,1989,1996,2007,2012) and O’Brien and Opie (2004,2006), along with other approaches including clustering (Shea, 2007) and mutual information (Azhar, 2017). My analysis of relatively simple facial recognition ANNs reveals that the structural relations between represented facial categories can vary between different ANNs and may reflect artificial relations rather than intuitive concepts of facial similarity. However, my analysis of a broad range of ANNs categorising various aspects of colour reveals that they develop robust and consistent task-dependent structural representations that do match the relational structure of corresponding human colour judgements. The task relevant structural arrangement of representations that are developed by these networks provides empirical support for the use of structural-resemblance approaches to explaining how ANNs represent and process information.