Under what circumstances do local codes emerge in feed-forward neural networks

Localist coding schemes are more easily interpretable than the distributed schemes but generally believed to be biologically implausible. Recent results have found highly selective units and object detectors in NNs that are indicative of local codes (LCs). Here we undertake a constructionist study on feed-forward NNs and find LCs emerging in response to invariant features, and this finding is robust until the invariant feature is perturbed by 40%. Decreasing the number of input data, increasing the relative weight of the invariant features and large values of dropout all increase the number of LCs. Longer training times increase the number of LCs and the turning point of the LC-epoch curve correlates well with the point at which NNs reach 90-100% on both test and training accuracy. Pseudo-deep networks (2 hidden layers) which have many LCs lose them when common aspects of deep-NN research are applied (large training data, ReLU activations, early stopping on training accuracy and softmax), suggesting that LCs may not be found in deep-NNs. Switching to more biologically feasible constraints (sigmoidal activation functions, longer training times, dropout, activation noise) increases the number of LCs. If LCs are not found in the feed-forward classification layers of modern deep-CNNs these data suggest this could either be caused by a lack of (moderately) invariant features being passed to the fully connected layers or due to the choice of training conditions and architecture. Should the interpretability and resilience to noise of LCs be required, this work suggests how to tune a NN so they emerge.