Steering Sparse Autoencoder Latents to Control Dynamic Head Pruning in Vision Transformers (Student Abstract)

TL;DR

This paper introduces a framework that uses Sparse Autoencoders to interpret and control dynamic head pruning in Vision Transformers, enhancing efficiency and interpretability.

Contribution

It presents a novel method integrating SAEs with ViT pruning to produce interpretable, class-specific head subsets that maintain accuracy while reducing head usage.

Findings

Per-class steering improves accuracy from 76% to 82%.

Head usage reduces from 0.72 to 0.33 with the proposed method.

Sparse latent features enable class-specific control of pruning.

Abstract

Dynamic head pruning in Vision Transformers (ViTs) improves efficiency by removing redundant attention heads, but existing pruning policies are often difficult to interpret and control. In this work, we propose a novel framework by integrating Sparse Autoencoders (SAEs) with dynamic pruning, leveraging their ability to disentangle dense embeddings into interpretable and controllable sparse latents. Specifically, we train an SAE on the final-layer residual embedding of the ViT and amplify the sparse latents with different strategies to alter pruning decisions. Among them, per-class steering reveals compact, class-specific head subsets that preserve accuracy. For example, bowl improves accuracy (76% to 82%) while reducing head usage (0.72 to 0.33) via heads h2 and h5. These results show that sparse latent features enable class-specific control of dynamic pruning, effectively bridging pruning efficiency and mechanistic interpretability in ViTs.