A Highly Parallel and Energy Efficient Three-Dimensional Multilayer CMOS-RRAM Accelerator for Tensorized Neural Network
IEEE TRANSACTIONS ON NANOTECHNOLOGY
Authors: Huang, Hantao; Ni, Leibin; Wang, Kanwen; Wang, Yuangang; Yu, Hao
Abstract
It is a grand challenge to develop highly parallel yet energy-efficient machine learning hardware accelerator. This paper introduces a three-dimensional (3-D) multilayer CMOS-RRAM accelerator for a tensorized neural network. Highly parallel matrix-vector multiplication can be performed with low power in the proposed 3-D multilayer CMOS-RRAM accelerator. The adoption of tensorization can significantly compress the weight matrix of a neural network using much fewer parameters. Simulation results using the benchmark MNIST show that the proposed accelerator has 1.283x speed-up, 4.276x energy-saving, and 9.339x area-saving compared to the 3-D CMOS-ASIC implementation; and 6.37x speed-up and 2612x energy-saving compared to 2-D CPU implementation. In addition, 14.85x model compression can be achieved by tensorization with acceptable accuracy loss.
Adaptive algorithms for low-rank and sparse matrix recovery with truncated nuclear norm
INTERNATIONAL JOURNAL OF MACHINE LEARNING AND CYBERNETICS
Authors: Qian, Wenchao; Cao, Feilong
Abstract
Recent studies have shown that the use of the truncated nuclear norm (TNN) in low-rank and sparse matrix decomposition (LRSD) can realize a better approximation to rank function of matrix, and achieve effectively recovery effects. This paper addresses the algorithms for LRSD with adaptive TNN (LRSD-ATNN), and designs an efficient algorithmic frame inspired by the alternating direction method of multiple (ADMM) and the accelerated proximal gradient approach (APG). To establish the adaptive algorithms, the method of singular value estimate is utilized to find adaptively the number of truncated singular value. Experimental results on synthetic data as well as real visual data show the superiority of the proposed algorithm in effectiveness in comparison with the state-of-the-art methods.
COA
Publication 
Loading ......
