arXivDaily arXiv每日学术速递 周一至周五更新
北美退伍军人节(Veterans Day),12 日无更新
arXiv 1803.06396cs.LGstat.ML
收藏

恢复和改进递归反向传播

Reviving and Improving Recurrent Back-Propagation

University of Toronto(多伦多大学);Uber ATG Toronto(优步先进技术集团多伦多);Vector Institute(向量研究所);Rice University(莱斯大学);Baylor College of Medicine(贝勒医学院);Canadian Institute for Advanced Research(加拿大高等研究院)

Renjie Liao, Yuwen Xiong, Ethan Fetaya, Lisa Zhang, KiJung Yoon, Xaq Pitkow, Raquel Urtasun, Richard Zemel

摘要:

在本文中,我们重新审视了递归反向传播(RBP)算法,讨论了它适用的条件以及如何在深度神经网络中满足这些条件。我们证明了RBP可能是不稳定的,并提出了两个变体的基础上共轭梯度的正规方程(CG-RBP)和Neumann系列(Neumann-RBP)。我们进一步研究了Neumann-RBP和时间反向传播(BPTT)及其截断版本(TBPTT)之间的关系。我们的Neumann-RBP具有与TBPTT相同的时间复杂度,但只需要恒定的内存,而TBPTT的内存成本与截断步骤的数量线性缩放。我们研究了所有RBP变体以及BPTT和TBPTT在三个不同的应用领域:连续Hopfield网络的联想记忆,引文网络中使用图神经网络的文档分类和全连接网络的超参数优化。所有的实验表明,RBP,特别是Neumann-RBP变体,是高效和有效的优化收敛递归神经网络。代码发布于:\url{https://github.com/lrjconan/RBP}。

英文摘要:

In this paper, we revisit the recurrent back-propagation (RBP) algorithm, discuss the conditions under which it applies as well as how to satisfy them in deep neural networks. We show that RBP can be unstable and propose two variants based on conjugate gradient on the normal equations (CG-RBP) and Neumann series (Neumann-RBP). We further investigate the relationship between Neumann-RBP and back propagation through time (BPTT) and its truncated version (TBPTT). Our Neumann-RBP has the same time complexity as TBPTT but only requires constant memory, whereas TBPTT's memory cost scales linearly with the number of truncation steps. We examine all RBP variants along with BPTT and TBPTT in three different application domains: associative memory with continuous Hopfield networks, document classification in citation networks using graph neural networks and hyperparameter optimization for fully connected networks. All experiments demonstrate that RBPs, especially the Neumann-RBP variant, are efficient and effective for optimizing convergent recurrent neural networks. Code is released at: \url{https://github.com/lrjconan/RBP}.

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。

↑