Linear Computation Coding for Convolutional Neural Networks

Müller R, Rosenberger H, Reichenbach M (2023)

Publication Type: Conference contribution

Publication year: 2023

DOI: 10.1109/SSP53291.2023.10207943

Abstract

Linear computation coding is a recent method aimed to reduce the computational effort for matrix-vector multiplications. Previous algorithms work especially well for large, dense matrices. These algorithms are adapted here for use in convolutional neural networks, where filter kernels typically consist of small, square matrices. For kernel matrices with a uniform element distribution, the number of additions (multiplications are counted as multiple additions) can be reduced compared to a standard implementation by a factor of two to five depending on the kernel size. For an implementation of LeNet5, the number of adders required is reduced by a factor of up to 2.74 compared to standard methods.

Authors with CRIS profile

Ralf Müller Professur für Informationsübertragung Hans Rosenberger Professur für Informationsübertragung

Involved external institutions

Brandenburgische Technische Universität Cottbus-Senftenberg (BTU) DE Germany (DE)

How to cite

APA:

Müller, R., Rosenberger, H., & Reichenbach, M. (2023). Linear Computation Coding for Convolutional Neural Networks.

MLA:

Müller, Ralf, Hans Rosenberger, and Marc Reichenbach. "Linear Computation Coding for Convolutional Neural Networks." 2023.

BibTeX: Download