将微调视为子图搜索，揭示模型学习动态新机制。
Fine-Tuning is Subgraph Search: A New Lens on Learning Dynamics
- 把模型看作有冗余的计算图，微调即在其中搜索最优子图。
- 实验验证该视角下学习过程更高效，且平衡任务性能与泛化能力。
- 适合研究模型内部机制或设计更优训练算法的研究者。
机制可解释性旨在逆向解析模型行为。尽管现有研究关注特定行为的静态机制，模型内部的学习动态仍待探索。本文提出一种基于内在维度的微调方法，将模型视为针对特定任务具有冗余的计算图，将微调过程视为在此图中搜索并优化子图的过程。基于此假设，我们提出电路微调（circuit-tuning）算法，通过迭代构建目标任务的子图，并以启发式方式更新相关参数。我们通过精心设计的实验验证了该假设，并详细分析了微调过程中的学习动态。随后，在更复杂的任务上进行实验，表明电路微调可在目标任务性能与通用能力之间取得良好平衡。本工作为微调动态提供了新分析视角，揭示了训练过程的内在机制，并启发了神经网络训练算法的改进设计。
原文摘要 · Abstract (English)
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 a fine-tuning method for analyzing the mechanism behind learning. Inspired by the concept of intrinsic dimension, we view a model as a computational graph with redundancy for a specific task, and treat the fine-tuning process as a search for and optimization of a subgraph within this graph. Based on this hypothesis, we propose circuit-tuning, an algorithm that iteratively builds the subgraph for a specific task and updates the relevant parameters in a heuristic way. We first validate our hypothesis through a carefully designed experiment and provide a detailed analysis of the learning dynamics during fine-tuning. Subsequently, we conduct experiments on more complex tasks, demonstrating that circuit-tuning could strike a balance between the performance on the target task and the general capabilities. Our work offers a new analytical method for the dynamics of fine-tuning, provides new findings on the mechanisms behind the training process, and inspires the design of superior algorithms for the training of neural networks.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。