Open AccessReduce the Computing Time for SpikeProp by Approximation of Spike Response Function
Author(s) -
Koya Kawanishi,
Haruhiko Takase,
Hiroharu Kawanaka,
Shinji Tsuruoka
Publication year - 2016
Publication title -
procedia computer science
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.334
H-Index - 76
ISSN - 1877-0509
DOI - 10.1016/j.procs.2016.08.162
Subject(s) - computer science , spike (software development) , function approximation , computation , generalization , approximation error , approximation algorithm , function (biology) , transformation (genetics) , simple (philosophy) , linear approximation , algorithm , polynomial , polynomial time approximation scheme , mathematical optimization , time complexity , artificial neural network , mathematics , artificial intelligence , nonlinear system , mathematical analysis , biochemistry , chemistry , philosophy , physics , software engineering , epistemology , quantum mechanics , evolutionary biology , gene , biology
pikeProp, which is proposed by Bohte, is a kind of spiking neural networks. Booij extends it to handle multiple spikes. As a result, SpikeProp can perform transformation of spike sequences. On the other hand, a problem is arose: it requires much computation time. In this article, we solve the problem by approximation: approximate the spike response function by a polynomial. For approximation, we take care of two issues: (1) minimize approximation error on the ascending slope, (2) smoothly decrease to zero. We show the effectiveness of the approximation by simple experiments. With the approximated spike response function, we reduced computation time of less than one tenth of the original model without degrading training ability and generalization ability
Accelerating Research
Robert Robinson Avenue,
Oxford Science Park, Oxford
OX4 4GP, United Kingdom
Address
John Eccles HouseRobert Robinson Avenue,
Oxford Science Park, Oxford
OX4 4GP, United Kingdom