LG-CAV: Train Any Concept Activation Vector with Language Guidance

TL;DR

Language-Guided CAV (LG-CAV) is proposed to harness the abundant concept knowledge within the certain pre-trained vision-language models to harness the abundant concept knowledge within the certain concept-based models (e.g., CLIP).

Abstract

Concept activation vector (CAV) has attracted broad research interest in explainable AI, by elegantly attributing model predictions to specific concepts. However, the training of CAV often necessitates a large number of high-quality images, which are expensive to curate and thus limited to a predefined set of concepts. To address this issue, we propose Language-Guided CAV (LG-CAV) to harness the abundant concept knowledge within the certain pre-trained vision-language models (e.g., CLIP). This method allows training any CAV without labeled data, by utilizing the corresponding concept descriptions as guidance. To bridge the gap between vision-language model and the target model, we calculate the activation values of concept descriptions on a common pool of images (probe images) with vision-language model and utilize them as language guidance to train the LG-CAV. Furthermore, after training high-quality LG-CAVs related to all the predicted classes in the target model, we propose the activation sample reweighting (ASR), serving as a model correction technique, to improve the performance of the target model in return. Experiments on four datasets across nine architectures demonstrate that LG-CAV achieves significantly superior quality to previous CAV methods given any concept, and our model correction method achieves state-of-the-art performance compared to existing concept-based methods. Our code is available at https://github.com/hqhQAQ/LG-CAV.

Figures (8)

• Figure 1: The quality of CAV is significantly affected by the number of training images. Here concept accuracy estimates whether the CAV faithfully represents its corresponding concept. Concept-to-class accuracy measures the similarity between the CAV and its strongly semantic-related class.
• Figure 2: (A) LG-CAV is trained guided by activations of concept descriptions on the probe images from VL model. (B) The distribution of activation values on a concept named "Skyscraper" (from the Broden dataset bau2017dissection) in the target model (ResNet18) and VL model (CLIP) differs a lot.
• Figure 3: Top: The original CAV is defined as the weight vector for its represented concept in the binary linear classifier. Bottom: The LG-CAV is learned by mimicking the activation values of its represented concept on the probe images $\mathcal{R}$ using VL model. Besides, three modules (GA module, CE module, and DSR module) are proposed to enhance the quality of LG-CAV.
• Figure 4: Ablation experiments on probe images (selection strategy & image number).
• Figure 5: (A) Activation values of the LG-CAV & (B) Model correction example.

Theorems & Definitions (1)

• proof