Train Sparse Autoencoders Efficiently by Utilizing Features Correlation
This work addresses efficiency issues in training sparse autoencoders for interpreting language models, which is an incremental improvement in a domain-specific area.
The authors tackled the computational challenge of training Sparse Autoencoders at scale by proposing KronSAE, which uses Kronecker product decomposition to reduce memory and computational overhead, and introduced mAND, a differentiable activation function that improves interpretability and performance.
Sparse Autoencoders (SAEs) have demonstrated significant promise in interpreting the hidden states of language models by decomposing them into interpretable latent directions. However, training SAEs at scale remains challenging, especially when large dictionary sizes are used. While decoders can leverage sparse-aware kernels for efficiency, encoders still require computationally intensive linear operations with large output dimensions. To address this, we propose KronSAE, a novel architecture that factorizes the latent representation via Kronecker product decomposition, drastically reducing memory and computational overhead. Furthermore, we introduce mAND, a differentiable activation function approximating the binary AND operation, which improves interpretability and performance in our factorized framework.