# Can Stability be Detrimental? Better Generalization through Gradient Descent Instabilities

arXiv: https://arxiv.org/abs/2412.17613 · HTML: https://mlsanity.com/paper/2412.17613/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 20 of its 29 references that are in the map.
- First posted on arXiv 23 Dec 2024.

## Builds on (most cited first, 20 of 20)

- [Deep Residual Learning for Image Recognition](https://mlsanity.com/paper/1512.03385/) (2015)
- [Very Deep Convolutional Networks for Large-Scale Image Recognition](https://mlsanity.com/paper/1409.1556/) (2014)
- [Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift](https://mlsanity.com/paper/1502.03167/) (2015)
- [Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms](https://mlsanity.com/paper/1708.07747/) (2017)
- [On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima](https://mlsanity.com/paper/1609.04836/) (2016)
- [Fantastic Generalization Measures and Where to Find Them](https://mlsanity.com/paper/1912.02178/) (2019)
- [Sharp Minima Can Generalize For Deep Nets](https://mlsanity.com/paper/1703.04933/) (2017)
- [Gradient Descent on Neural Networks Typically Occurs at the Edge of Stability](https://mlsanity.com/paper/2103.00065/) (2021)
- [ASAM: Adaptive Sharpness-Aware Minimization for Scale-Invariant Learning of Deep Neural Networks](https://mlsanity.com/paper/2102.11600/) (2021)
- [Train longer, generalize better: closing the generalization gap in large batch training of neural networks](https://mlsanity.com/paper/1705.08741/) (2017)
- [Three Factors Influencing Minima in SGD](https://mlsanity.com/paper/1711.04623/) (2017)
- [The large learning rate phase of deep learning: the catapult mechanism](https://mlsanity.com/paper/2003.02218/) (2020)
- [Self-Stabilization: The Implicit Bias of Gradient Descent at the Edge of Stability](https://mlsanity.com/paper/2209.15594/) (2022)
- [Understanding Gradient Descent on Edge of Stability in Deep Learning](https://mlsanity.com/paper/2205.09745/) (2022)
- [Implicit Bias of SGD for Diagonal Linear Networks: a Provable Benefit of Stochasticity](https://mlsanity.com/paper/2106.09524/) (2021)
- [A Modern Look at the Relationship between Sharpness and Generalization](https://mlsanity.com/paper/2302.07011/) (2023)
- [Measurements of Three-Level Hierarchical Structure in the Outliers in the Spectrum of Deepnet Hessians](https://mlsanity.com/paper/1901.08244/) (2019)
- [SGD with Large Step Sizes Learns Sparse Features](https://mlsanity.com/paper/2210.05337/) (2022)
- [On the Maximum Hessian Eigenvalue and Generalization](https://mlsanity.com/paper/2206.10654/) (2022)
- [Learning Rates as a Function of Batch Size: A Random Matrix Theory Approach to Neural Network Training](https://mlsanity.com/paper/2006.09092/) (2020)

## Abstract (arXiv)

Traditional analyses of gradient descent optimization show that, when the largest eigenvalue of the loss Hessian - often referred to as the sharpness - is below a critical learning-rate threshold, then training is 'stable' and training loss decreases monotonically. Recent studies, however, have suggested that the majority of modern deep neural networks achieve good performance despite operating outside this stable regime. In this work, we demonstrate that such instabilities, induced by large learning rates, move model parameters toward flatter regions of the loss landscape. Our crucial insight lies in noting that, during these instabilities, the orientation of the Hessian eigenvectors rotate. This, we conjecture, allows the model to explore regions of the loss landscape that display more desirable geometrical properties for generalization, such as flatness. These rotations are a consequence of network depth, and we prove that for any network with depth > 1, unstable growth in parameters cause rotations in the principal components of the Hessian, which promote exploration of the parameter space away from unstable directions. Our empirical studies reveal an implicit regularization effect in gradient descent with large learning rates operating beyond the stability threshold. We find these lead to excellent generalization performance on modern benchmark datasets.