Attribution Patching Outperforms Automated Circuit Discovery

Autor:	Syed, Aaquib, Rager, Can, Conmy, Arthur
Rok vydání:	2023
Předmět:	Computer Science - Machine Learning Computer Science - Artificial Intelligence Computer Science - Computation and Language
Druh dokumentu:	Working Paper
Popis:	Automated interpretability research has recently attracted attention as a potential research direction that could scale explanations of neural network behavior to large models. Existing automated circuit discovery work applies activation patching to identify subnetworks responsible for solving specific tasks (circuits). In this work, we show that a simple method based on attribution patching outperforms all existing methods while requiring just two forward passes and a backward pass. We apply a linear approximation to activation patching to estimate the importance of each edge in the computational subgraph. Using this approximation, we prune the least important edges of the network. We survey the performance and limitations of this method, finding that averaged over all tasks our method has greater AUC from circuit recovery than other methods. Comment: 6 main paper pages, 6 additional pages. NeurIPS 2023 ATTRIB Workshop
Databáze:	arXiv
Externí odkaz:	http://arxiv.org/abs/2310.10348 Zobrazit plný text záznamu View this record from Arxiv