Ensembling Sparse Autoencoders
This work addresses the problem of limited feature extraction in SAEs for researchers and practitioners in interpretable AI, though it is incremental as it applies existing ensembling techniques to SAEs.
The paper tackles the limitation that single sparse autoencoders (SAEs) capture only a subset of features from neural network activations by proposing to ensemble multiple SAEs using bagging and boosting methods. The result shows improved reconstruction of activations, feature diversity, stability, and better performance on downstream tasks like concept detection and spurious correlation removal.
Sparse autoencoders (SAEs) are used to decompose neural network activations into human-interpretable features. Typically, features learned by a single SAE are used for downstream applications. However, it has recently been shown that SAEs trained with different initial weights can learn different features, demonstrating that a single SAE captures only a limited subset of features that can be extracted from the activation space. Motivated by this limitation, we propose to ensemble multiple SAEs through naive bagging and boosting. Specifically, SAEs trained with different weight initializations are ensembled in naive bagging, whereas SAEs sequentially trained to minimize the residual error are ensembled in boosting. We evaluate our ensemble approaches with three settings of language models and SAE architectures. Our empirical results demonstrate that ensembling SAEs can improve the reconstruction of language model activations, diversity of features, and SAE stability. Furthermore, ensembling SAEs performs better than applying a single SAE on downstream tasks such as concept detection and spurious correlation removal, showing improved practical utility.