REVIEW 7 cited by
Exploring Visual Interpretability for Contrastive Language-Image Pre-training
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
Contrastive Language-Image Pre-training (CLIP) learns rich representations via readily available supervision of natural language. It improves the performance of downstream vision tasks, including but not limited to the zero-shot, long tail, segmentation, retrieval, caption, and video. However, the visual explainability of CLIP is rarely studied, especially for the raw feature map. To provide visual explanations of its predictions, we propose the Image-Text Similarity Map (ITSM). Based on it, we surprisingly find that CLIP prefers the background regions than the foregrounds, and shows erroneous visualization results against human understanding. This phenomenon is universal for both vision transformers and convolutional networks, which suggests this problem is unique and not owing to certain network. Experimentally, we find the devil is in the pooling part, where inappropriate pooling methods lead to a phenomenon called semantic shift. For this problem, we propose the Explainable Contrastive Language-Image Pre-training (ECLIP), which corrects the explainability via the Masked Max Pooling. Specifically, to avoid the semantic shift, we replace the original attention pooling by max pooling to focus on the confident foreground, with guidance from free attention during training. Experiments on three datasets suggest that ECLIP greatly improves the explainability of CLIP, and beyond previous explainability methods at large margins. The code will be released later.
Forward citations
Cited by 7 Pith papers
-
AnchorPrune: Relevance-Anchored Contextual Expansion for Visual Token Pruning
A two-stage training-free visual token pruning method protects query-critical tokens before expanding with informative context, preserving 97.6% of full-token performance at 94.4% compression on LLaVA-NeXT-7B.
-
AnchorPrune: Relevance-Anchored Contextual Expansion for Visual Token Pruning
AnchorPrune prunes visual tokens by first selecting a protected query-relevance anchor and then greedily adding important, non-redundant context, preserving up to 97.6% of full-token accuracy with only 160 of 2,880 tokens.
-
What's in an Earth Embedding? An Explainability Analysis of Location Encoders
Location embeddings from geographic INRs can be decomposed into sparse latent concepts, natural language concepts, and visual features while retaining high reconstruction capability.
-
Right Predictions, Misleading Explanations: On the Vulnerability of Vision-Language Model Explanations
X-Shift is a grey-box attack that perturbs patch-level visual features in VLMs to shift explanation heatmaps without changing the predicted output.
-
Right Predictions, Misleading Explanations: On the Vulnerability of Vision-Language Model Explanations
The paper shows that explanation heatmaps in vision-language models can be redirected to irrelevant image regions via imperceptible patch perturbations without changing model predictions, using a new attack called X-Shift.
-
Grad-ECLIP: Gradient-based Visual and Textual Explanations for CLIP
Grad-ECLIP produces gradient-based visual and textual explanation heatmaps for CLIP by applying channel and spatial weights to token features instead of relying on sparse self-attention maps.
-
AnchorPrune: Relevance-Anchored Contextual Expansion for Visual Token Pruning
Protecting an adaptive query-relevance anchor before importance-weighted contextual expansion preserves up to 97.6% of full-token VLM accuracy with only 160 of 2,880 visual tokens.
Discussion (0). Sign in to comment.