arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2012.08795cs.LGcs.AI
收藏

基于二阶梯度的神经网络大批量训练研究

Study on the Large Batch Size Training of Neural Networks Based on the Second Order Gradient

Institute of Microelectronics of the Chinese Academy of Sciences(中国科学院微电子研究所)

Fengli Gao, Huicai Zhong

摘要:

深度神经网络(DNN)中的大批量训练具有众所周知的“泛化间隙”,该间隙显著地导致泛化性能下降。然而,目前还不清楚不同的批量大小如何影响NN的结构。在这里,我们将理论与实验相结合,探索基本结构特性的演变,包括梯度,参数更新步长,和损失更新步长的神经网络在不同的批量大小。我们提供了新的指导,以提高泛化,这是进一步验证了两个设计的方法,涉及丢弃小损失样本和调度批量大小。提出了一种基于曲率的学习率(CBLR)算法,以更好地拟合神经网络各层的曲率变化(影响大批量训练的敏感因素)。作为CBLR的一种近似,中值曲率LR算法(MLR)被发现获得与分层自适应速率缩放(LARS)算法相当的性能。我们的理论结果和算法提供了基于几何的解释现有的研究。此外,我们证明了层明智的LR算法,例如LARS,可以被视为CBLR的特殊情况。最后,我们推导出大批量训练的理论几何图像,并表明所有网络参数都倾向于集中在其相关的最小值。

英文摘要:

Large batch size training in deep neural networks (DNNs) possesses a well-known 'generalization gap' that remarkably induces generalization performance degradation. However, it remains unclear how varying batch size affects the structure of a NN. Here, we combine theory with experiments to explore the evolution of the basic structural properties, including gradient, parameter update step length, and loss update step length of NNs under varying batch sizes. We provide new guidance to improve generalization, which is further verified by two designed methods involving discarding small-loss samples and scheduling batch size. A curvature-based learning rate (CBLR) algorithm is proposed to better fit the curvature variation, a sensitive factor affecting large batch size training, across layers in a NN. As an approximation of CBLR, the median-curvature LR (MCLR) algorithm is found to gain comparable performance to Layer-wise Adaptive Rate Scaling (LARS) algorithm. Our theoretical results and algorithm offer geometry-based explanations to the existing studies. Furthermore, we demonstrate that the layer wise LR algorithms, for example LARS, can be regarded as special instances of CBLR. Finally, we deduce a theoretical geometric picture of large batch size training, and show that all the network parameters tend to center on their related minima.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑