Improving Sparse Autoencoder with Dynamic Attention
This work addresses the sparsity-reconstruction trade-off in SAEs for interpretability of foundation models, offering a more flexible activation function that improves both metrics.
Sparse autoencoders (SAEs) for interpreting foundation models suffer from a trade-off between sparsity and reconstruction quality. The authors propose a cross-attention SAE with sparsemax activation that dynamically adjusts sparsity per neuron, achieving lower reconstruction loss and high-quality concepts in top-n classification tasks.
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.