Proceedings of the Annual Meeting of the Cognitive Science Society

To understand the representations learned by neural networks(NNs), various methods of measuring unit selectivity havebeen developed. Here we undertake a comparison of four suchmeasures on AlexNet: localist selectivity (Bowers et al., 2014);precision (Zhou et al., 2015); class-conditional mean activityselectivity CCMAS (Morcos et al., 2018); and top-class se-lectivity. In contrast with previous work on recurrent neuralnetworks (RNNs), we fail to find any 100% selective ‘local-ist units’ in AlexNet, and demonstrate that the precision andCCMAS measures are misleading and suggest a much higherlevel of selectivity than is warranted. We also generated ac-tivation maximization (AM) images that maximally activatedindividual units and found that under (5%) of units in fc6 andconv5 produced interpretable images of objects, whereas fc8produced over 50% interpretable images. Furthermore, theinterpretable images in the hidden layers were not associatedwith highly selective units. We also consider why localist rep-resentations are learned in RNNs and not AlexNet.