Pro-MA: Progressively Margin-based Attribution in Pre-trained Vision-Language Models

Abstract Knowledge attribution is focused on analyzing the internal knowledge architecture of neural networks, aiming to accurately identify the factual knowledge stored within a model. Existing knowledge attribution methods concentrate on Pre-trained Language Models (PLMs), with limited exploration in Pre-trained Vision-Language Models (VLMs). However, the knowledge attribution methods of PLMs cannot be directly utilized on VLMs due to different decision-making processes and unstable integral gradients. Specifically, VLMs involve extra knowledge-irrelevant steps like localization and show high variance in gradient distributions along the attribution paths. To address the above challenges, we propose Progressive Margin-based Attribution (Pro-MA) to achieve more accurate attribute knowledge attribution in VLMs. Firstly, we employ cross-attribute comparison to identify and eliminate irrelevant knowledge by constructing high-similarity negative samples. Secondly, we utilize progressive attribution scoring that quantifies the contribution of neurons to the output by the margin of the attribute-value loss function along the path of active neurons. Evaluation on the OVAD and VAW datasets shows that Pro-MA outperforms the state-of-the-art knowledge attribution methods.