Attribution Analysis Meets Model Editing: Advancing Knowledge Correction in Vision Language Models with VisEdit

2024-08-19

arXiv.org

AI Summary

This study addresses the critical bottleneck in vision-language model (VLM) knowledge correction—namely, the necessity of full retraining. It is the first to systematically identify the pivotal role of mid-to-late-layer visual representations and prompt-correlated regions in factual prediction. Building on this insight, we propose VisEdit, the first interpretability-driven, training-free editing framework for VLMs. VisEdit enables precise cross-modal intervention via three core mechanisms: contribution allocation analysis, noise-perturbation attribution, and intermediate visual feature localization. Evaluated across multiple mainstream VLM backbones and established public editing benchmarks, VisEdit consistently outperforms LLM-adapted editors, achieving an average 12.6% improvement in editing accuracy. Moreover, it demonstrates strong generalization and minimal side effects. This work establishes a novel paradigm for efficient, interpretable, and parameter-efficient knowledge editing in vision-language models.

Technology Category

• Computer Vision: Large Vision Models
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Model editing aims to correct outdated or erroneous knowledge in large models without costly retraining. Recent research discovered that the mid-layer representation of the subject's final token in a prompt has a strong influence on factual predictions, and developed Large Language Model (LLM) editing techniques based on this observation. However, for Vision-LLMs (VLLMs), how visual representations impact the predictions from a decoder-only language model remains largely unexplored. To the best of our knowledge, model editing for VLLMs has not been extensively studied in the literature. In this work, we employ the contribution allocation and noise perturbation methods to measure the contributions of visual representations for token predictions. Our attribution analysis shows that visual representations in mid-to-later layers that are highly relevant to the prompt contribute significantly to predictions. Based on these insights, we propose VisEdit, a novel model editor for VLLMs that effectively corrects knowledge by editing intermediate visual representations in regions important to the edit prompt. We evaluated VisEdit using multiple VLLM backbones and public VLLM editing benchmark datasets. The results show the superiority of VisEdit over the strong baselines adapted from existing state-of-the-art editors for LLMs.

Problem

• Visual Language Models
• Knowledge Editing
• Prediction Mechanism

Innovation

• VisEdit
• Visual Language Model Editing
• Knowledge Accuracy Improvement