Simulating Training Spikes of Deep Neural Networks by Layer-wise Random Feature Models
cs.LG
Submitted: 2026-10-09
Updated: 2026-10-09
Sources
- High dimensional analysis reveals conservative sharpening and a stochastic edge of stability
- Layer Normalization
- Communication-Efficient Language Model Training Scales Reliably and Robustly: Scaling Laws for DiLoCo
- Adaptive Gradient Methods at the Edge of Stability
- Logarithmic-time Schedules for Scaling Language Models with Momentum
- Training Compute-Optimal Large Language Models
- The large learning rate phase of deep learning: the catapult mechanism
- RoBERTa: A Robustly Optimized BERT Pretraining Approach
- A Solvable Model of Neural Scaling Laws
- To Clip or not to Clip: the Dynamics of SGD with Gradient Clipping in High-Dimensions
- A Theory on Adam Instability in Large-Scale Machine Learning
- On Layer Normalization in the Transformer Architecture
- Catapults in SGD: spikes in the training loss and their impact on generalization through feature learning
Related papers
- Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
- AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
- Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
- Convergence issues in Relational Concept Analysis based on AOC-posets
- Beliefs Beyond Posteriors: The Predictive Cost of ELBO Binding
- Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks