Stochastic Training is Not Necessary for Generalization
This work addresses the problem of understanding generalization mechanisms in deep learning for researchers, suggesting that the perceived difficulty of full-batch training may be due to optimization properties and community focus on tuning for small-batch methods, making it incremental by questioning a widely held assumption.
The paper challenges the belief that stochastic gradient descent (SGD) is necessary for generalization in neural networks by showing that non-stochastic full-batch training can achieve comparable performance to SGD on CIFAR-10 with modern architectures, using explicit regularization to replace SGD's implicit regularization.
It is widely believed that the implicit regularization of SGD is fundamental to the impressive generalization behavior we observe in neural networks. In this work, we demonstrate that non-stochastic full-batch training can achieve comparably strong performance to SGD on CIFAR-10 using modern architectures. To this end, we show that the implicit regularization of SGD can be completely replaced with explicit regularization even when comparing against a strong and well-researched baseline. Our observations indicate that the perceived difficulty of full-batch training may be the result of its optimization properties and the disproportionate time and effort spent by the ML community tuning optimizers and hyperparameters for small-batch training.