CLAICVFeb 26, 2024

Cross-Modal Projection in Multimodal LLMs Doesn't Really Project Visual Attributes to Textual Space

Georgia Tech
arXiv:2402.16832v237 citationsh-index: 13Has CodeACL
Originality Synthesis-oriented
AI Analysis

This work addresses the interpretability and design of MLLMs for researchers and practitioners, revealing that fine-tuning updates do not enhance projection networks, which is incremental as it clarifies existing architectures without proposing new methods.

The study investigated the role of cross-modal projection networks in multimodal large language models (MLLMs) during fine-tuning for domain-specific tasks like dermatology and agriculture, finding that domain-specific visual attributes are modeled by the large language model rather than the projection network, with experiments on 4 datasets and 2 fine-tuning settings showing no extraction of relevant attributes by the projection.

Multimodal large language models (MLLMs) like LLaVA and GPT-4(V) enable general-purpose conversations about images with the language modality. As off-the-shelf MLLMs may have limited capabilities on images from domains like dermatology and agriculture, they must be fine-tuned to unlock domain-specific applications. The prevalent architecture of current open-source MLLMs comprises two major modules: an image-language (cross-modal) projection network and a large language model. It is desirable to understand the roles of these two modules in modeling domain-specific visual attributes to inform the design of future models and streamline the interpretability efforts on the current models. To this end, via experiments on 4 datasets and under 2 fine-tuning settings, we find that as the MLLM is fine-tuned, it indeed gains domain-specific visual capabilities, but the updates do not lead to the projection extracting relevant domain-specific visual attributes. Our results indicate that the domain-specific visual attributes are modeled by the LLM, even when only the projection is fine-tuned. Through this study, we offer a potential reinterpretation of the role of cross-modal projections in MLLM architectures. Project webpage: https://claws-lab.github.io/projection-in-MLLMs/

Code Implementations1 repo
Foundations

The foundational work for this paper's niche, ranked by how specifically the neighbourhood builds on it — not by global fame.

Your Notes