Improving Sparse Autoencoder with Dynamic Attention

2026-04-16

AI Summary

This work addresses the challenge in sparse autoencoders where a fixed sparsity level often compromises the trade-off between reconstruction fidelity and interpretability. To overcome this limitation, the authors propose a dynamic sparse autoencoder based on cross-attention, wherein latent features serve as queries and a learnable dictionary functions as key-value pairs. The method introduces, for the first time, a sparsemax-based dynamic sparse attention mechanism that enables the model to adaptively determine the number of active neurons according to the input, without requiring additional regularization or meticulous hyperparameter tuning. Experimental results demonstrate that the proposed approach significantly reduces reconstruction loss while preserving high interpretability, and yields higher-quality concept representations in top-n classification tasks.

Technology Category

• Machine Learning: Deep Generative Models & Autoencoders
• Computer Vision: Representation Learning for Vision
• Cognitive Modeling & Cognitive Systems: Neural Spike Coding

Application Category

• Search and Retrieval-Augmented AI: Web query analysis, representation and understanding
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Recently, sparse autoencoders (SAEs) have emerged as a promising technique for interpreting activations in foundation models by disentangling features into a sparse set of concepts. However, identifying the optimal level of sparsity for each neuron remains challenging in practice: excessive sparsity can lead to poor reconstruction, whereas insufficient sparsity may harm interpretability. While existing activation functions such as ReLU and TopK provide certain sparsity guarantees, they typically require additional sparsity regularization or cherry-picked hyperparameters. We show in this paper that dynamically sparse attention mechanisms using sparsemax can bridge this trade-off, due to their ability to determine the activation numbers in a data-dependent manner. Specifically, we first explore a new class of SAEs based on the cross-attention architecture with the latent features as queries and the learnable dictionary as the key and value matrices. To encourage sparse pattern learning, we employ a sparsemax-based attention strategy that automatically infers a sparse set of elements according to the complexity of each neuron, resulting in a more flexible and general activation function. Through comprehensive evaluation and visualization, we show that our approach successfully achieves lower reconstruction loss while producing high-quality concepts, particularly in top-n classification tasks.

Problem

• sparse autoencoder
• sparsity
• interpretability
• reconstruction loss
• neuron activation

Innovation

• Sparse Autoencoder
• Dynamic Attention
• Sparsemax
• Cross-Attention Architecture
• Interpretability