Circuit-tuning: A Mechanistic Approach for Identifying Parameter Redundancy and Fine-tuning Neural Networks

Abstract: The study of mechanistic interpretability aims to reverse-engineer a model to explain its behaviors. While recent studies have focused on the static mechanism of a certain behavior, the learning dynamics inside a model remain to be explored. In this work, we develop an interpretable fine-tuning method for analyzing the mechanism behind learning. We first introduce the concept of node-level intrinsic dimensionality to describe the learning process of a model in a computational graph. Based on our theory, we propose circuit-tuning, a two-stage algorithm that iteratively builds the minimal subgraph for a specific task and updates the key parameters in a heuristic way. Experimental results confirm the existence of the intrinsic dimensionality at the node level and demonstrate the effectiveness of our method for transparent and interpretable fine-tuning. We visualize and analyze the circuits before, during, and after fine-tuning, providing new insights into the self-organization mechanism of a neural network in the learning process.