Sparse Autoencoders (SAEs)
Sparse Autoencoders (SAEs)
Auto-generated stub. Edit this file to add more details.
An interpretability technique using overcomplete dictionary learning to decompose neural network activations into sparse latent features.
Why It Matters
Used here as the primary method to probe for representational superposition in time-series models, demonstrating its utility in cross-domain mechanistic analysis.
Evidence
applying sparse autoencoders (SAEs), a tool from mechanistic interpretability, to probe the internal representations of PatchTST.
Related Papers
Metadata & Links
- created_at
- 2026-05-09T05:56:39Z