MMNeuron: Discovering Neuron-Level Domain-Specific Interpretation in Multimodal Large Language Model
Huo, Jiahao, Yan, Yibo, Hu, Boren, Yue, Yutao, Hu, Xuming
–arXiv.org Artificial Intelligence
Projecting visual features into word embedding space has become a significant fusion strategy adopted by Multimodal Large Language Models (MLLMs). However, its internal mechanisms have yet to be explored. Inspired by multilingual research, we identify domain-specific neurons in multimodal large language models. Specifically, we investigate the distribution of domain-specific neurons and the mechanism of how MLLMs process features from diverse domains. Furthermore, we propose a three-stage framework for language model modules in MLLMs when handling projected image features, and verify this hypothesis using logit lens. Extensive experiments indicate that while current MLLMs exhibit Visual Question Answering (VQA) capability, they may not fully utilize domain-specific information. Manipulating domain-specific neurons properly will result in a 10\% change of accuracy at most, shedding light on the development of cross-domain, all-encompassing MLLMs in the future. Our code will be released upon paper notification.
arXiv.org Artificial Intelligence
Jun-16-2024
- Country:
- North America > United States
- New York (0.04)
- California (0.04)
- Washington > King County
- Seattle (0.04)
- Europe
- Middle East > Malta
- Eastern Region > Northern Harbour District > St. Julian's (0.04)
- Ireland > Leinster
- County Dublin > Dublin (0.04)
- Middle East > Malta
- Asia
- Singapore (0.04)
- Middle East > UAE
- Abu Dhabi Emirate > Abu Dhabi (0.04)
- China
- Hong Kong (0.04)
- Guangdong Province > Guangzhou (0.04)
- North America > United States
- Genre:
- Research Report (0.82)
- Industry:
- Health & Medicine (0.68)
- Technology: