Gradient Smoothing: Coupling Layer-wise Updates for Improved Optimization
梯度平滑:耦合逐层更新以改进优化
机构 * Vector Institute(向量研究所)
专题命中 预训练与数据 :LLM(abstract_cn);language model(abstract);pretraining(abstract);post-training(abstract)
AI总结 提出深度梯度增强框架,通过窗口平滑算子沿深度方向变换优化器更新,在不改变模型架构或训练目标下提升多种架构和任务的优化与泛化性能。
Comments Published in the Proceedings of the 43rd International Conference on Machine Learning (ICML 2026)