Train One Sparse Autoencoder Across Multiple Sparsity Budgets to Preserve Interpretability and Accuracy
This addresses a computational and flexibility bottleneck for researchers and practitioners using SAEs for neural network interpretability, offering an incremental improvement over existing methods.
The paper tackles the problem of sparse autoencoders (SAEs) requiring separate models for different sparsity levels, which increases computational costs, by introducing a HierarchicalTopK training objective that trains a single SAE to optimize reconstructions across multiple sparsity levels simultaneously. Experiments on Gemma-2 2B show it achieves Pareto-optimal trade-offs between sparsity and explained variance, outperforming traditional SAEs and preserving high interpretability scores at higher sparsity.
Sparse Autoencoders (SAEs) have proven to be powerful tools for interpreting neural networks by decomposing hidden representations into disentangled, interpretable features via sparsity constraints. However, conventional SAEs are constrained by the fixed sparsity level chosen during training; meeting different sparsity requirements therefore demands separate models and increases the computational footprint during both training and evaluation. We introduce a novel training objective, \emph{HierarchicalTopK}, which trains a single SAE to optimise reconstructions across multiple sparsity levels simultaneously. Experiments with Gemma-2 2B demonstrate that our approach achieves Pareto-optimal trade-offs between sparsity and explained variance, outperforming traditional SAEs trained at individual sparsity levels. Further analysis shows that HierarchicalTopK preserves high interpretability scores even at higher sparsity. The proposed objective thus closes an important gap between flexibility and interpretability in SAE design.