Reproduce “A Closer Look at the Explainability of Contrastive Language-Image Pre-training” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yi Li, Hualiang Wang, Yiqun Duan, Jiheng Zhang, Xiaomeng Li
This paper systematically analyzes and resolves key explainability issues in raw Contrastive Language-Image Pre-training (CLIP) models, such as 'opposite visualization' and 'noisy activations'. It introduces "CLIP Surgery," a zero-training method that modifies the inference architecture and features to produce highly accurate Class Attention Maps (CAMs), subsequently enabling enhanced performance on open-vocabulary tasks like semantic segmentation and multi-label recognition without additional fine-tuning.
Key insights
- CLIP's self-attention layers, with heterologous parameters, can cause 'opposite visualization' by linking inconsistent semantic regions, while Feed-Forward Networks (FFNs) further push features away from positive labels.
- Noisy activations in CLIP explanations stem from redundant features among categories, with activations resembling those from an 'empty string' input.
- Enhancing model explainability without fine-tuning can directly improve performance on various open-vocabulary downstream tasks, such as semantic segmentation and multi-label recognition.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.