ETRI-Knowledge Sharing Plaform

KOREAN
논문 검색
Type SCI
Year ~ Keyword

Detail

Conference Paper Data Compression Hardware of the ReLu Output in Convolution Neural Network
Cited - time in scopus Download 11 time Share share facebook twitter linkedin kakaostory
Authors
Juyeob Kim, Miyoung Lee, Byoungjo Kim, Jinkyu Kim, Sungmin Kim, Juehyun Lee
Issue Date
2017-11
Citation
International Conference on Future Generation Communication and Networking (FGCN) 2017 (ASTL 146), pp.1-7
Language
English
Type
Conference Paper
Project Code
17HS1600, Feasibility Study of Blue IT based on Human Body Research, Cho Il Yeon
Abstract
This paper describes a compression method of ReLu layer output to reduce the bandwidth of external memory through a dedicated hardware implementation. The output of the ReLu layer is suitable for variable length coding as its property. This paper proposes a hardware encoder through Golomb-Rice coding with analysis of various lossless compression algorithms in terms of compression efficiency and hardware implementation. Through the proposed hardware, we explain the chunk-based compression method and explain how to achieve maximum 1/3 compression in chunk units.
KSP Keywords
Compression Algorithm, Compression method, Convolution neural network(CNN), Golomb-Rice coding, Hardware Implementation, Variable length coding, compression efficiency, data compression, dedicated hardware, external memory, lossless compression