Attribution Graph
A graph structure used to analyze interactions among features in neural networks, commonly employed in mechanistic interpretability research to predict the effect of ablation or perturbation on model outputs.
Timeline (last 90 days)
Sparse autoencoder (SAE) feature attribution maps are actually determined by the feature's "margin", i.e., the distance between the pre-activation value and the TopK truncation threshold.
In a write operation that scales feature encodings by three times, the graph built from previously active features has a median relative error of at least 0.247, while the maximum error of the margin graph is only 0.033.
All Facts (2)
Sparse autoencoder (SAE) feature attribution maps are actually determined by the feature's "margin", i.e., the distance between the pre-activation value and the TopK truncation threshold.
50%UnverifiedIn a write operation that scales feature encodings by three times, the graph built from previously active features has a median relative error of at least 0.247, while the maximum error of the margin graph is only 0.033.
50%