揭示Transformer注意力学习中的聚焦与稀释循环机制
Focus and Dilution: The Multi-stage Learning Process of Attention
- 通过梯度流分析，发现注意力学习存在周期性聚焦-稀释过程
- 注意力逐步聚焦高频词，再通过嵌入扰动实现焦点稀释
- 适用于理解Transformer训练动态，适合研究模型内部机制者
基于Transformer的模型在多个领域取得显著成功，但其训练动态仍不清晰。本文在单层Transformer和马尔可夫数据设置下，通过梯度流分析揭示注意力学习中反复出现的聚焦-稀释循环，并证明该循环可分解为若干阶段：首先，嵌入与投影快速凝聚为秩一结构，注意力参数基本冻结；随后注意力参数增长，驱动对高频词的频率导向聚焦；随着注意力演化，嵌入产生高阶扰动，引发质量重分布机制，逐步稀释聚焦；最后，低频词间的微小不对称性打破退化临界点，开启新的嵌入方向，启动下一周期。在合成马尔可夫数据以及WikiText、TinyStories上的实验验证了预测的阶段和周期性动态。
原文摘要 · Abstract (English)
Transformer-based models have achieved remarkable success across a wide range of domains, yet our understanding of their training dynamics remains limited. In this work, we identify a recurrent focus-dilution cycle in attention learning and provide a rigorous explanation in a one-layer Transformer setting for Markovian data via gradient-flow analysis. Using stage-wise linearization around critical points, we show that a single focus-dilution cycle can be decomposed into a sequence of distinct stages. First, embedding and projection rapidly condense to a rank-one structure, while attention parameters remain effectively frozen. Then, the attention parameters begin to increase, inducing a frequency-driven focus toward high-frequency tokens. As attention continues to evolve, it generates next-order perturbations in embeddings, leading to a mass-redistribution mechanism that progressively dilutes this focus. Finally, small asymmetries among low-frequency tokens lift a degenerate critical point, opening new embedding directions and initiating the next cycle. Experiments on synthetic Markovian data as well as WikiText and TinyStories corroborate the predicted stages and cyclical dynamics.
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。