• Search Research Projects
  • Search Researchers
  • How to Use
  1. Back to previous page

Neural Network Engine based on a Blind Learning Algorithm

Research Project

Project/Area Number 20K04626
Research Category

Grant-in-Aid for Scientific Research (C)

Allocation TypeMulti-year Fund
Section一般
Review Section Basic Section 21060:Electron device and electronic equipment-related
Research InstitutionWaseda University

Principal Investigator

OHSAWA Takashi  早稲田大学, 理工学術院(情報生産システム研究科・センター), 教授(任期付) (10613391)

Project Period (FY) 2020-04-01 – 2023-03-31
Project Status Completed (Fiscal Year 2022)
Budget Amount *help
¥4,420,000 (Direct Cost: ¥3,400,000、Indirect Cost: ¥1,020,000)
Fiscal Year 2022: ¥1,170,000 (Direct Cost: ¥900,000、Indirect Cost: ¥270,000)
Fiscal Year 2021: ¥2,340,000 (Direct Cost: ¥1,800,000、Indirect Cost: ¥540,000)
Fiscal Year 2020: ¥910,000 (Direct Cost: ¥700,000、Indirect Cost: ¥210,000)
Keywordsディープ・ニューラル・ネットワーク / イン・メモリ・コンピューティング / 不揮発性メモリ / オフセット電圧 / オペアンプ / トレーニング・プログラム / MOSFET / 閾値電圧ばらつき / ニューラルネットワーク / 不揮発性メモリデバイス / アンデュアランス / ニューロモルフィックシステム / 新機能メモリ
Outline of Research at the Start

本研究は不揮発性の新機能メモリを用いて人工ニューラルネットワークにおける推論を高速かつローパワーで実行するエンジンにおいて、その学習をオンチップで行えるようにするものである。一般的に確立された学習アルゴリズムとして逆誤差伝搬法があるが、これに基づいた学習をオンチップで実行するにはアナログの乗算回路が必要となるなど、シナプス回路に新機能メモリ素子を用いて小型化したメリットの意味が薄れてしまう問題が出てくる。そこで、逆誤差伝搬法よりもハードウェアへの負担が軽いくなる新しい学習アルゴリズムを提案し、それに基づいた学習が実行できるニューラルネットワークエンジンの設計を可能とする技術を研究する。

Outline of Final Research Achievements

We have been studying on training methods to recover the accuracy degradation in an inference accelerator of neural networks due to the circuit imperfection. Although it was initially supposed that "on-chip training" would be the best method for achieving the goal, it was found that it introduced additional factors which degrade the accuracy because the circuits which are implemented in the backward path and are not used in the forward inference path lead to another root cause of the accuracy degradation. "In-situ training" was another candidate for the solution. However, this suffers from the nonvolatile memory devices' endurance issues, non-linearity and asymmetry switching characteristics issues. Finally, we proposed a novel training method named "hardware-conscious software training" in which the circuits' imperfections are implemented into the training program. It was verified that the inference accuracy which was degraded by the circuit parameter fluctuations can be recovered.

Academic Significance and Societal Importance of the Research Achievements

ニューラルネットワークによる画像認識などを不揮発性メモリを用いた専用集積回路で実現することは、汎用コンピュータにおけるソフトウェアで実現する方式に比べて消費電力と計算速度の点で大きなメリットがある。しかし、専用集積回路チップを製造する際に避けられない素子特性のばらつきにより、認識精度が低下しまう問題があった。我々は、この問題を解決する新たな学習方法を提案した。これは、素子特性のばらつきの一部を測定した後に、それらを学習プログラムに取り込む方法であり、認識精度劣化がほぼ完全に回復できることを示せた。この方法は不揮発性メモリの特性劣化やスイッチングの非線形性・非対称性の課題も解決できるものである。

Report

(4 results)
  • 2022 Annual Research Report   Final Research Report ( PDF )
  • 2021 Research-status Report
  • 2020 Research-status Report
  • Research Products

    (11 results)

All 2023 2022 2021 2020

All Journal Article (8 results) (of which Int'l Joint Research: 5 results,  Peer Reviewed: 8 results) Presentation (3 results) (of which Int'l Joint Research: 3 results)

  • [Journal Article] A Fully Analog Deep Neural Network Inference Accelerator with Pipeline Registers Based on Master-Slave Switched Capacitors2023

    • Author(s)
      Yaxin MEI, Takashi OHSAWA
    • Journal Title

      IEICE Trans. Electron.

      Volume: vol. E106-C, no.9

    • Related Report
      2022 Annual Research Report
    • Peer Reviewed / Int'l Joint Research
  • [Journal Article] A gate leakage current-powered loadless 4T SRAM with immunity against random dopant fluctuation and surface roughness in silicon-silicon dioxide interface2023

    • Author(s)
      Yihan ZHU, Takashi OHSAWA
    • Journal Title

      Jpn. J. Appl. Phys.

      Volume: 62 Issue: SC Pages: SC1004-SC1004

    • DOI

      10.35848/1347-4065/aca33b

    • Related Report
      2022 Annual Research Report
    • Peer Reviewed / Int'l Joint Research
  • [Journal Article] A Binarized Spiking Neural Network Based on Auto-Reset LIF Neurons and Large Signal Synapses Using STT-MTJs2023

    • Author(s)
      Haoyan LIU, Takashi OHSAWA
    • Journal Title

      Jpn. J. Appl. Phys.

      Volume: 62 Issue: 4 Pages: 044501-044501

    • DOI

      10.35848/1347-4065/acc9f4

    • Related Report
      2022 Annual Research Report
    • Peer Reviewed / Int'l Joint Research
  • [Journal Article] A Low-Cost Training Method of ReRAM Inference Accelerator Chips for Binarized Neural Networks to Recover Accuracy Degradation due to Statistical Variabilities2022

    • Author(s)
      Zian CHEN, Takashi OHSAWA
    • Journal Title

      IEICE Transactions on Electronics

      Volume: E105.C Issue: 8 Pages: 375-384

    • DOI

      10.1587/transele.2021ECP5040

    • NAID

      130008149058

    • ISSN
      0916-8524, 1745-1353
    • Year and Date
      2022-08-01
    • Related Report
      2022 Annual Research Report 2021 Research-status Report
    • Peer Reviewed / Int'l Joint Research
  • [Journal Article] A Loadless 4T SRAM Powered by Gate Leakage Current with a High Tolerance for Fluctuations in Device Parameters2022

    • Author(s)
      Yihan ZHU, Takashi OHSAWA
    • Journal Title

      Jpn. J. Appl. Phys.

      Volume: 61 Issue: SC Pages: SC1053-SC1053

    • DOI

      10.35848/1347-4065/ac44ce

    • Related Report
      2022 Annual Research Report
    • Peer Reviewed / Int'l Joint Research
  • [Journal Article] Compact Model of Magnetic Tunnel Junctions for SPICE Simulation Based on Switching Probability2021

    • Author(s)
      Haoyan LIU, Takashi OHSAWA
    • Journal Title

      IEICE Transactions on Electronics

      Volume: E104.C Issue: 3 Pages: 121-127

    • DOI

      10.1587/transele.2020ECP5011

    • NAID

      130007993130

    • ISSN
      0916-8524, 1745-1353
    • Year and Date
      2021-03-01
    • Related Report
      2020 Research-status Report
    • Peer Reviewed
  • [Journal Article] Array Design of High-Density Emerging Memories Making Clamped Bit-Line Sense Amplifier Compatible with Dummy Cell Average Read Scheme2020

    • Author(s)
      Ziyue ZHANG, Takashi OHSAWA
    • Journal Title

      IEICE Transactions on Electronics

      Volume: E103.C Issue: 8 Pages: 372-380

    • DOI

      10.1587/transele.2019ECP5039

    • NAID

      130007883640

    • ISSN
      0916-8524, 1745-1353
    • Year and Date
      2020-08-01
    • Related Report
      2020 Research-status Report
    • Peer Reviewed
  • [Journal Article] Co-Design of Binary Processing in Memory ReRAM Array andDNN Model Optimization Algorithm2020

    • Author(s)
      Yue GUAN, Takashi OHSAWA
    • Journal Title

      IEICE Transactions on Electronics

      Volume: E103-C NO. 11 Pages: 685-692

    • Related Report
      2020 Research-status Report
    • Peer Reviewed
  • [Presentation] Spin-Transfer Torque Magnetic Tunnel Junction Model Based on Fokker-Planck Equation for Stochastic Circuit Simulations2022

    • Author(s)
      Haoyan LIU, Takashi OHSAWA
    • Organizer
      22nd IEEE International Conference on Nanotechnology
    • Related Report
      2022 Annual Research Report
    • Int'l Joint Research
  • [Presentation] A Loadless 4T SRAM Cell Powered by Gate Leakage Current and Tolerant of Random Dopant Fluctuation and Surface Roughness at Si-SiO2 Interface2022

    • Author(s)
      Yihan ZHU, Takashi OHSAWA
    • Organizer
      2022 International Conference on Solid State Devices and Materials (SSDM)
    • Related Report
      2022 Annual Research Report
    • Int'l Joint Research
  • [Presentation] Spin-Transfer Torque Magnetic Tunnel Junction Model Based on Fokker-Planck Equation for Stochastic Circuit Simulations2022

    • Author(s)
      Haoyan Liu, Takashi Ohsawa
    • Organizer
      22nd IEEE International Conference on Nanotechnology
    • Related Report
      2021 Research-status Report
    • Int'l Joint Research

URL: 

Published: 2020-04-28   Modified: 2024-01-30  

Information User Guide FAQ News Terms of Use Attribution of KAKENHI

Powered by NII kakenhi