finding
active
finding:jacobian-matching-removes-the-memorization-feature-in-repeated-data-example-leaving-constituent-features-slightly-rotatedJacobian matching removes the memorization feature in repeated-data example, leaving constituent features slightly rotated
Toy-model fix of unfaithfulness.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Positive result with caveats.
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.
- Adding Jacobian matching loss dramatically reduces datapoint-feature memorization metricfinding0.856Result of Jacobian matching.
- Motivates Jacobian connection.
- Penalizing difference between transcoder and MLP Jacobians, computed cheaply via random vector products.
- Limit on Jacobian matching.
- Transcoders could cheat Jacobian matching via large-weight, very negative-bias features exploiting the L1/L0 gaphypothesis0.768Gradient masking risk; mitigated by tanh L1 or penalizing Jacobian contribution L1.
- Larger rotations between shared features skew decoder cosine similarities downward in the toy modelfinding0.751Supports 'same features, used differently' interpretation.
- Limitation.
- Key quote connecting path redundancy to interferometric information encoding.