A Design of Fully Stochastic Computing Neurons Focused on the Gain of Sigmoid Functions

DOI Open Access

Bibliographic Information

Other Title
  • シグモイド関数のゲインに着目した完全ストカスティック計算ニューロンの設計

Search this article

Description

Stochastic computing (SC) is an approximate computation with probabilities. SC-based neural networks (NNs) recently draw attention as an efficient implementation of NNs. For fully SC neurons, in which all the calculation of the neurons are implemented by only SC, it is one of the major challenges to reduce the calculation error of the neurons and to maintain the recognition accuracy of NNs. In this paper, we focus on the gain of sigmoid functions, which are used as activation functions of neurons, and propose two-types of fully SC neurons with a selective type multiply-accumulate circuit [nagura] and a linear finite state machine (linear FSM). We also propose a learning algorithm for the proposed NNs to reduce the calculation error of the neurons. Experimental results show that, compared with the previous SC-based NNs, the proposed NNs can achieve the high recognition accuracy.

Journal

Related Projects

See more

Keywords

Details 詳細情報について

  • CRID
    1390288535620486400
  • DOI
    10.14923/transinfj.2020fip0008
  • ISSN
    18810225
    18804535
  • Text Lang
    ja
  • Article Type
    journal article
  • Data Source
    • JaLC
    • KAKEN
  • Abstract License Flag
    Disallowed

Report a problem

Back to top