在视觉语言模型中保留局部化补丁语义
Logit Lens Supervision for Patch-Level Explanations in Vision-Language Models
AI总结:
为解决视觉语言模型中图像与文本 tokens 信息扩散问题,提出 Logit Lens Loss(LLL)以保留图像 tokens 的局部化语义,提升模型可解释性和分割任务性能。
摘要:
Logit Lens 已被提出用于可视化对 LLM 回答贡献最大的 tokens。最近,Logit Lens 也被证明适用于自回归的视觉语言模型(VLMs),其中它以热图的形式展示图像 tokens 的概念内容,例如在给定图像中哪些图像 tokens 可能描绘猫的概念。然而,图像 tokens 的视觉内容往往扩散到语言 tokens,从而导致视觉信息的局部性大部分被破坏,这使得 Logit Lens 可视化在可解释性上不可用。为了解决这个问题,我们引入了一个互补的损失函数到下一个 token 预测(NTP)中,以防止视觉 tokens 丢失从相应图像补丁继承的视觉表示。所提出的 Logit Lens 损失(LLL)旨在使视觉 token 嵌入更语义上与描述其图像区域的文本概念对齐(例如,包含猫的补丁与单词“cat”)。无需任何架构修改或大规模训练。这样,LLL 会约束自注意力层中图像和文本 tokens 的混合,以防止图像 tokens 丢失其局部化的视觉信息。正如我们的实验所显示的,LLL 不仅使 Logit Lens 实用,通过在图像中产生有意义的对象置信度图,还提高了在以视觉为中心的任务如分割上的性能,而无需附加任何特殊头。
英文摘要:
Modern autoregressive Vision-Language Models (VLMs) can generate fluent answers while their visual-token representations become weakly tied to the image regions from which they originate. This limits patch-level explainability: a visual token should remain interpretable as the image patch it represents. We study this issue through Logit Lens maps, obtained by projecting each visual-token embedding through the LLM vocabulary head to measure how strongly each image patch is associated with a queried textual concept, such as "cat". We introduce Logit Lens Loss (LLL), a lightweight auxiliary objective that preserves localized visual semantics by directly aligning object-related visual tokens with the vocabulary concepts describing their image regions. LLL requires no architectural modification, mask decoder, or large-scale retraining. We evaluate LLL through both explanation quality and downstream performance. Across LLaVA-v1.5-7B and Qwen2.5-VL-7B, LLL yields sharper object confidence maps, improves grounding and hallucination metrics, transfers to zero-shot pointing, and preserves VQA performance.
机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。