method
active
method:tanh-l1-regularizationtanh L1 regularization
Sparsity regularizer approximating L0 used to avoid duplicate features.
Neighborhood — ranked by edge-count
Papers (1)
paper
Methods (1)
method
- TranscodersusesDecomposition method for activations; VPD is compared against transcoders in sparsity-reconstruction tradeoff.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Sparsity penalty weighting by sum of per-layer decoder norms, enabling comparability with per-layer SAEs.
- Removes superposition at cost of loss.
- Alternative penalty that spreads features across layers and finds only shared features.
- Preliminary experiment.
- SAE variant trained on ResNet18.
- KL penalty to base model so self-explanation stays about the same model.
- Architecture compared.
- L1-of-norms is preferable for loss comparability and exposing layer- or model-specific featuresclaim0.674All experiments used L1-of-norms.