Multimodal Neurons in Pretrained Text-Only Transformers
This work addresses the interpretability of multimodal generalization in AI models, providing insights into how neurons bridge modalities, which is incremental but offers specific mechanistic understanding.
The paper tackled the problem of understanding how pretrained text-only transformers generalize to vision tasks, finding that translation between modalities occurs deeper within the transformer through 'multimodal neurons' that convert visual representations into text, with experiments showing these neurons operate on specific visual concepts and have a causal effect on image captioning.
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.