paper-with-me

Papers

The Lifecycle of the Spectral Edge: From Gradient Learning to Weight-Decay Compression

2026-04-08 · Yongzhong Xu arxiv

We decompose the spectral edge -- the dominant direction of the Gram matrix of parameter updates -- into its gradient and weight-decay components during grokking in two sequence tasks (Dyck-1 and SCAN). We find a sharp two-phase lifecycle: before grokking the edge is gradient-driven and functionally active; at grokking, gradient and weight decay align, and the edge becomes a compression axis that is perturbation-flat yet ablation-critical (>4000x more impactful than random directions). Three universality classes emerge (functional, mixed, compression), predicted by the gap flow equation. Nonlinear probes show information is re-encoded, not lost (MLP $R^2=0.99$ where linear $R^2=0.86$), and removing weight decay post-grok reverses compression while preserving the algorithm.

📄 PDF Abstract BibTeX arXiv:2604.07380

Code (0)

등록된 구현이 없습니다.

Similar Papers 제목 키워드 기반

Low-Rank Decay for Grokking in Scale-Invariant Transformers: A Spectral-Geometric View

2026-06-03 · Mingyu Li arxiv

Modern Transformer architectures frequently employ normalization mechanisms such as RMSNorm and Query-Key Normalization, making parts of the model approximately scale-invariant with respect to weight magnitudes. In this …

AlphaDecay: Module-wise Weight Decay for Heavy-Tailed Balancing in LLMs

2025-06-17 · Di He, Ajay Jaiswal, Songjun Tu, Li Shen 외

Weight decay is a standard regularization technique for training large language models (LLMs). While it is common to assign a uniform decay rate to every layer, this approach overlooks the structural diversity of LLMs an…

Diversity

Initialization and Regularization of Factorized Neural Layers

2021-05-03 · ICLR 2021 1 · Mikhail Khodak, Neil Tenenholtz, Lester Mackey, Nicolò Fusi

Factorized layers--operations parameterized by products of two or more matrices--occur in a variety of deep learning contexts, including compressed model training, certain types of knowledge distillation, and multi-head …

Knowledge DistillationModel CompressionTensor DecompositionUnsupervised Pre-training

Learning to Forget: Continual Learning with Adaptive Weight Decay

2026-04-29 · Aditya A. Ramesh, Alex Lewandowski, Jürgen Schmidhuber arxiv

Continual learning agents with finite capacity must balance acquiring new knowledge with retaining the old. This requires controlled forgetting of knowledge that is no longer needed, freeing up capacity to learn. Weight …

Continual Learning

On the Overlooked Pitfalls of Weight Decay and How to Mitigate Them: A Gradient-Norm Perspective

2020-11-23 · NeurIPS 2023 11 · Zeke Xie, Zhiqiang Xu, Jingzhao Zhang, Issei Sato 외

Weight decay is a simple yet powerful regularization technique that has been very widely used in training of deep neural networks (DNNs). While weight decay has attracted much attention, previous studies fail to discover…