Data for: 3360696

First, we reason formally through the effect of dropout and training stochasticity on gradient descent. Then, we conduct classification experiments on the ImageNet data set, as well as regression experiments in the OneNow Reinforcement Learning data set.

A network layer's weight matrix is quantified via Singular Value Decomposition and Conditioning ratios. Our regression network appeared to be well conditioned. However, we find that learning for large-scale classification applications is likely to be capped by poor conditioning.