Theory of Deep Learning IIb: Optimization Properties of SGD
arXiv:1801.02254
Abstract
In Theory IIb we characterize with a mix of theory and experiments the optimization of deep convolutional networks by Stochastic Gradient Descent. The main new result in this paper is theoretical and experimental evidence for the following conjecture about SGD: SGD concentrates in probability -- like the classical Langevin equation -- on large volume, "flat" minima, selecting flat minimizers which are with very high probability also global minimizers
Cited by in corpus (10)
- Multi-Objective Loss Balancing for Physics-Informed Deep Learning
- Theory of Deep Learning III: explaining the non-overfitting puzzle
- Strength of Minibatch Noise in SGD
- Consistent Sparse Deep Learning: Theory and Computation
- On the Distributional Properties of Adaptive Gradients
- Noise and Fluctuation of Finite Learning Rate Stochastic Gradient Descent
- What training reveals about neural network complexity
- Global Convergence of SGD On Two Layer Neural Nets
- On the Bias-Variance Tradeoff: Textbooks Need an Update
- HydaLearn: Highly Dynamic Task Weighting for Multi-task Learning with Auxiliary Tasks