Multimodal Neurons in Pretrained Text-Only Transformers

TL;DR

This paper investigates how individual neurons in a pretrained text transformer can be responsible for converting visual information into language, revealing the neural basis of multimodal understanding.

Contribution

It introduces a method to identify and analyze 'multimodal neurons' in text transformers that integrate visual concepts into language generation.

Findings

Multimodal neurons operate on specific visual concepts.

These neurons have a causal effect on image captioning.

Conversion between modalities occurs deeper within the transformer.

Abstract

Language models demonstrate remarkable capacity to generalize representations learned in one modality to downstream tasks in other modalities. Can we trace this ability to individual neurons? We study the case where a frozen text transformer is augmented with vision using a self-supervised visual encoder and a single linear projection learned on an image-to-text task. Outputs of the projection layer are not immediately decodable into language describing image content; instead, we find that translation between modalities occurs deeper within the transformer. We introduce a procedure for identifying "multimodal neurons" that convert visual representations into corresponding text, and decoding the concepts they inject into the model's residual stream. In a series of experiments, we show that multimodal neurons operate on specific visual concepts across inputs, and have a systematic causal effect on image captioning.