Abstract:

In this work, we present a methodology that integrates mechanistic and gradient-based interpretability techniques to reduce bias in the CLIP Vision Transformer, focusing on enhancing inner alignment. We begin by defining the CLIP model and reviewing the role of Sparse Autoencoders (SAEs) in mechanistic interpretability, specifically highlighting the use of the Jumped SAE configuration. We then introduce Grad ECLIP—a gradient-based method tailored for CLIP—to visualize the model’s focus areas. We apply our approach to detect and mitigate the reification bias associated with sexual objectification in CLIP, wherein the model disproportionately focuses on certain visual features (specifically, the bust region in images of semi-nude women). Using the Sexual OBjectification and EMotion (SOBEM) dataset of 280 images (140 images of semi-nude women and 140 images of the same women wearing a shirt), we extract residual activations from layers 22, 23, and 24 of the CLIP Vision Transformer, train nine different SAEs with expansion factors of x8, x16, and x32, and select the most appropriate SAE based on statistical evaluation. By integrating the selected SAE into CLIP and analyzing the resulting Grad ECLIP heatmaps—evaluated via Intersection over Union (IoU) with manually created bust masks—we test whether our intervention successfully reduces the bias. Finally, an ablation strategy is applied to validate the causal impact of our intervention on bias reduction and to improve the model’s inner alignment.

Link of implementation: https://github.com/MysticDeepAI/HeatMapSae-A-Gradient-Centric-Approach-to-Explaining-CLIP-Through-Sparse-Latent-Activations

1. Introduction

1.1. Presentation of the CLIP Model and Its Relevance in Multimodal Systems

The CLIP (Contrastive Language–Image Pre-training) model (Radford et al)represents a significant advancement in multimodal systems by jointly embedding images and textual descriptions into a shared latent space. Developed to bridge the gap between visual and textual data, CLIP utilizes a Vision Transformer (ViT) architecture, which has shown remarkable performance in various computer vision tasks. Its ability to understand and relate concepts across modalities makes it a powerful tool in contexts ranging from image retrieval to content moderation. However, the complexity and high dimensionality of its internal representation...

Posts

Wikitag Contributions

Comments