(R)NNs too expressive?: No relationship between neural architectures' theoretical expressivity and their empirical...

This work investigates why recurrent neural networks (RNNs) tend to learn phonological patterns that are unattested or dispreferred by humans. Specifically, we explore the hypothesis that their over-generation is caused by their excess expressive capacity - they are beyond the limited complexity class that contains the set of attested phonological patterns. We compared these over-expressive RNNs against the weaker convolutional neural networks (CNNs) on a battery of string recognition tasks. We find that the expressivity of a model's architecture does not predict the string classes that it excels at recognizing. Instead, we suggest that CNNs' position-invariant biases better explain their successes in our experiment.