Interpretable Textual Neuron Representations for NLP

TL;DR

This paper adapts input optimization techniques from computer vision to NLP, creating interpretable neuron representations that reveal syntactic differences in language models.

Contribution

It introduces a method using gradient ascent with Gumbel softmax to generate n-gram representations that better activate target neurons in NLP models.

Findings

Gradient ascent with Gumbel softmax outperforms naive corpus search.

Reveals differences in syntax awareness between language and visual models.

Provides interpretable neuron representations for NLP.

Abstract

Input optimization methods, such as Google Deep Dream, create interpretable representations of neurons for computer vision DNNs. We propose and evaluate ways of transferring this technology to NLP. Our results suggest that gradient ascent with a gumbel softmax layer produces n-gram representations that outperform naive corpus search in terms of target neuron activation. The representations highlight differences in syntax awareness between the language and visual models of the Imaginet architecture.