Rethinking Sparse Autoencoders: Select-and-Project for Fairness and Control from Encoder Features Alone
It addresses fairness and control in AI models, particularly for debiasing visual language models, with incremental improvements over existing methods.
The paper tackles the problem of debiasing sparse autoencoder representations by challenging the assumption that features are in decoder weights, introducing an encoder-focused method that improves fairness metrics by up to 3.2 times and advances state-of-the-art VLM debiasing by up to 1.8 times while preserving performance.
Sparse Autoencoders (SAEs) have proven valuable due to their ability to provide interpretable and steerable representations. Current debiasing methods based on SAEs manipulate these sparse activations presuming that feature representations are housed within decoder weights. We challenge this fundamental assumption and introduce an encoder-focused alternative for representation debiasing, contributing three key findings: (i) we highlight an unconventional SAE feature selection strategy, (ii) we propose a novel SAE debiasing methodology that orthogonalizes input embeddings against encoder weights, and (iii) we establish a performance-preserving mechanism during debiasing through encoder weight interpolation. Our Selection and Projection framework, termed S\&P TopK, surpasses conventional SAE usage in fairness metrics by a factor of up to 3.2 and advances state-of-the-art test-time VLM debiasing results by a factor of up to 1.8 while maintaining downstream performance.