Neural networks can understand compositional functions that humans do not, in the context of emergent communication

We show that it is possible to craft transformations that, applied to\ncompositional grammars, result in grammars that neural networks can learn\neasily, but humans do not. This could explain the disconnect between current\nmetrics of compositionality, that are arguably human-centric, and the ability\nof neural networks to generalize to unseen examples. We propose to use the\ntransformations as a benchmark, ICY, which could be used to measure aspects of\nthe compositional inductive bias of networks, and to search for networks with\nsimilar compositional inductive biases to humans. As an example of this\napproach, we propose a hierarchical model, HU-RNN, which shows an inductive\nbias towards position-independent, word-like groups of tokens.\n