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

TL;DR

This paper demonstrates that neural networks can understand certain compositional functions that humans do not, using crafted transformations and a new benchmark to measure and compare inductive biases related to compositionality.

Contribution

It introduces transformations that create a disconnect between human and neural network understanding of compositional grammars, and proposes ICY as a benchmark to evaluate neural inductive biases.

Findings

Neural networks can learn transformed compositional grammars easily.

The ICY benchmark measures neural inductive biases towards compositionality.

HU-RNN shows an inductive bias towards position-independent, word-like token groups.

Abstract

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.