arXiv:2609.34091 · v2

From Perturbation Sensitivity to Sparse Autoencoder Feature Interpretation: An EEG Foundation-Model Case Study

首次提交 2026-09-28 09:31

版本更新 2026-11-03 19:44

雷达观测 2026-11-04 23:39

摘要

Sparse autoencoders (SAEs) split a model's internal activity into many sparse units, or latents, and a latent is often named after a concept when removing that concept from the input changes its activity. We identify a consequential confound in this inference: if removing the concept changes the input more than the comparison edit does, a larger response may reflect a larger edit rather than the concept. Brain-signal (electroencephalography, EEG) models provide a controlled test, because how much an edit changes the EEG input can be measured directly. We measure the SAE response as the change in how often each latent is active, averaged over all latents. Across 27 combinations of model, dataset, and layer, removing the 8--12 Hz alpha band changes this response 7.33 times as much as removing an equally wide 30--34 Hz control band, but it also changes the signal far more. Matching the two edits per trial on squared time-domain distortion lowers the mean ratio to 1.42, with 19 of 27 combinations above one. Matching thus substantially reduces, but does not eliminate, the alpha--control contrast. The residual supports relative sensitivity under this single control, not general spectral specificity or an individual alpha-representing latent. We use an evidence-to-claim map to separate what this comparison supports from stronger claims of specificity and feature semantics.

作者备注：Revised version with expanded experimental analysis and discussion of limitations