Content and cluster analysis: assessing representational similarity in neural systems

If connectionism is to be an adequate theory of mind, we must have a theory representation for neural networks that allows for individual differences in weighting and architecture while preserving sameness, or at least simila rity, of content. In this paper we propose a procedure for measuring samene ss of content of neural representations. We argue that the correct way to c ompare neural representations is through analysis of the distances between neural activations, and we present a method for doing so. We then use the t echnique to demonstrate empirically that different artificial neural networ ks trained by backpropagation on the same categorization task, even with di fferent representational encodings of the input patterns and different numb ers of hidden units, reach states in which representations at the hidden un its are similar. We discuss how this work provides a rebuttal to Fodor and Lepore's critique of Paul Churchland's state space semantics.