.png){kind=icon}
返回
COMPUTING THE GRADIENT IN OPTIMIZATION ALGORITHMS FOR THE CP DECOMPOSITION IN CONSTANT MEMORY THROUGH TENSOR BLOCKING
.png){kind=icon}
OA
AI
DOI：10.1137/14097968X.png)
.png){kind=icon}
.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
The construction of the gradient of the objective function in gradient-based optimization algorithms for computing an r-term CANDECOMP/PARAFAC (CP) decomposition of an unstructured dense tensor is a key computational kernel. The best technique for efficiently implementing this operation has a memory consumption that scales linearly with the number of terms r and sublinearly with the number of elements of the tensor. We consider a blockwise computation of the CP gradient, reducing the memory requirements to a constant. This reduction is achieved by a novel technique that we call implicit block unfoldings, which combines the benefits of the block tensor unfoldings by [Ragnarsson and Van Loan, SIAM J. Matrix Anal. Appl., 33 (2012), pp. 149169] and the implicit unfoldings of [Phan, Tichavsky, and Cichocki, IEEE Trans. Signal Process., 61 (2013), pp. 4834-4846]. A heuristic algorithm for automatically choosing the division into subtensors is part of the proposed algorithm. The throughput that can be attained is essentially determined by the performance of a matrix product of two small matrices of constant size. Numerical experiments illustrate that the proposed method can outperform the current state-of-the-art by up to two orders of magnitude for large dense tensors in terms of memory consumption, while the increase of the execution time is no more than 5%. The proposed algorithm attained upward of 90% of the theoretical peak performance of the computer system, using no more than 50MB of memory, irrespective of the size of the tensor and the number of terms r.
Keyword:
CANDECOMP/PARAFAC
tensor rank decomposition
CP decomposition
CP gradient
implicit block unfolding
AI总结
对已上传原文的论文进行重点信息的提取，主要内容包括：简要概述、研究摘要、背景介绍、关键亮点、图文解析、展望与总结。
期刊
IF:
2.6
论文数:
5.1K
被引数:
1.8W
机构
引用论文
Additions and corrections: Potential energy models for nonbonding interactions in solid chlorine
PREAI
分享收藏A Fractional-Order Memristive Two-Neuron-Based Hopfield Neuron Network: Dynamical Analysis and Application for Image Encryption.png)
Mathematics
OAAI
MathematicsIF0
OAAI分享收藏Synthesis of high-performance parallel programs for a class of Ab Initio quantum chemistry models一类从头算量子化学模型的高性能并行程序的合成PROCEEDINGS OF THE IEEE
PROCEEDINGS OF THE IEEEIF25.9
PREAI
分享收藏Structural, optical and magnetic properties of Co-doped ZnO nanorods with hidden secondary phases
PREAI
分享收藏Enzymology with a Spin-Labeled Phospholipase C: Soluble Substrate Binding by 31P NMR from 0.005 to 11.7 T具有自旋标记的磷脂酶C的酶学: 通过 31P NMR从0.005到11.7 T的可溶性底物结合BiochemistryOAAI
BiochemistryIF0
OAAI分享收藏
查看更多内容