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 (6)
- Theory of Deep Learning III: explaining the non-overfitting puzzle
- Consistent Sparse Deep Learning: Theory and Computation
- On the Distributional Properties of Adaptive Gradients
- On the Bias-Variance Tradeoff: Textbooks Need an Update
- What training reveals about neural network complexity
- HydaLearn: Highly Dynamic Task Weighting for Multi-task Learning with Auxiliary Tasks