论文信息 - Memo No . 072 December 27 , 2017 Theory of Deep Learning IIb : Optimization Properties of SGD by

Memo No . 072 December 27 , 2017 Theory of Deep Learning IIb : Optimization Properties of SGD by

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. This work was supported by the Center for Brains, Minds and Machines (CBMM), funded by NSF STC award CCF 123 1216. This memo reports those parts of CBMM Memo 067 that are focused on optimization. The main reason is to straighten up the titles of the theory trilogy.

[1] Noah Golowich,et al. Musings on Deep Learning: Properties of SGD , 2017 .

[2] Tomaso A. Poggio,et al. Theory II: Landscape of the Empirical Risk in Deep Learning , 2017, ArXiv.

[3] Razvan Pascanu,et al. Sharp Minima Can Generalize For Deep Nets , 2017, ICML.

[4] Lorenzo Rosasco,et al. Why and when can deep-but not shallow-networks avoid the curse of dimensionality: A review , 2016, International Journal of Automation and Computing.

[5] Jorge Nocedal,et al. On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima , 2016, ICLR.

[6] Neil Genzlinger. A. and Q , 2006 .

[7] John N. Tsitsiklis,et al. Gradient Convergence in Gradient methods with Errors , 1999, SIAM J. Optim..

[8] John N. Tsitsiklis,et al. Neuro-Dynamic Programming , 1996, Encyclopedia of Machine Learning.

[9] S. Mitter,et al. Recursive stochastic algorithms for global optimization in R d , 1991 .

[10] B. Gidas. Global optimization via the Langevin equation , 1985, 1985 24th IEEE Conference on Decision and Control.