Neural networks can understand compositional functions that humans do not, in the context of emergent communication
This addresses the problem of aligning neural network inductive biases with human-like compositionality for researchers in AI and cognitive science, though it is incremental in proposing a new benchmark and model.
The paper demonstrates that neural networks can learn compositional functions from transformed grammars that humans cannot, revealing a disconnect between human-centric compositionality metrics and neural network generalization. It introduces the ICY benchmark to measure compositional inductive biases and proposes HU-RNN as a hierarchical model with a bias toward position-independent token groups.
We show that it is possible to craft transformations that, applied to compositional grammars, result in grammars that neural networks can learn easily, but humans do not. This could explain the disconnect between current metrics of compositionality, that are arguably human-centric, and the ability of neural networks to generalize to unseen examples. We propose to use the transformations as a benchmark, ICY, which could be used to measure aspects of the compositional inductive bias of networks, and to search for networks with similar compositional inductive biases to humans. As an example of this approach, we propose a hierarchical model, HU-RNN, which shows an inductive bias towards position-independent, word-like groups of tokens.