Top > Search of International Patents > FIXED-WEIGHTING-CODE LEARNING DEVICE

FIXED-WEIGHTING-CODE LEARNING DEVICE

Foreign code F180009562
File No. S2017-0527-C0
Posted date Nov 5, 2018
Country WIPO
International application number 2018JP004786
International publication number WO 2018168293
Date of international filing Feb 13, 2018
Date of international publication Sep 20, 2018
Priority data
  • P2017-048421 (Mar 14, 2017) JP
Title FIXED-WEIGHTING-CODE LEARNING DEVICE
Abstract A neural network circuit is provided with which it is possible to significantly reduce the area occupied by the connection unit of an FC-type neural network circuit. An analog-type neural network circuit SS constituting a learning device having a self-learning function and corresponding to a brain function, wherein the neural network SS comprises: a plurality (n) of input-side neurons NR; a plurality (m, and including cases when n=m) of output-side neurons NR; (n × m) connection units CN each connecting one input-side neuron NR and one output-side neuron NR; and a self-learning control unit C, the (n × m) connection units CN being constituted from connection units CN corresponding to only the positive weighting function as a brain function, and connection units CN corresponding to only the negative weighting function as the brain function.
Outline of related art and contending technology BACKGROUND ART
In recent years, the functions of the brain of the person using a deep neural network circuit (a so-called deep learning function) associated with a learning function of the research and development has been carried out.At this time, the neural network circuit in the case in specific implementation, a digital circuit is used and a case where an analog circuit is used and a case where.Here, the former has a high processing capability and a large-scale hardware is needed and the power consumption is increased, for example used in the data center or the like.While the latter, the processing capability is inferior as compared with the case of the digital circuit is used, as a hardware configuration and minimized to reduce the power consumption can be expected, for example a terminal connected to the data center or the like is often used for a device.Then, as for the latter prior art, the following technique disclosed in Patent Document 1 and the like.
Patent Document 1 is, configured by analog circuits (that is an analog type) neural network circuit, the output side of all the neurons of the input side and all of the neurons, corresponding to the respective weight of the resistance value by a connecting portion having the one-to-one connection, a so-called deep learning of the neural network circuit is coupled to all disclosed (FC(Full Connection) ) type.Weighting is in this case, the neural network circuit should correspond to the weighting and corresponding to the brain function, disclosed in Patent Document 1 is provided with a neural network circuit, the resistance value of each connection section is constituted by a variable resistance element correspond to each of the weight.FC-type neural network circuit such as, for example in the form of a neural network in the vicinity of the other coupled circuit, a large scale neural network than can be used as part of the circuit, itself further improves the processing capability, reduction in circuit scale (area) is desired.
Scope of claims (In Japanese)請求の範囲
[請求項1]
 脳機能に対応したアナログ型のニューラルネットワーク回路を備えた自己学習型の重み符号固定学習装置において、
 前記ニューラルネットワーク回路は、
 入力データに相当する入力信号がそれぞれ入力される複数(n個)の入力部と、
 出力データに相当する出力信号がそれぞれ出力される複数(m個且つn=mの場合を含む)の出力部と、
 一の前記入力部と一の前記出力部とをそれぞれ接続する(n×m)個の接続部と、
 により構成されており、
 前記出力データを前記出力部から前記重み符号固定学習装置に再入力した結果として前記入力部から出力されるデータが元の前記入力データと一致するように当該重み符号固定学習装置を制御して、前記自己学習の機能を実行させる制御手段を備え、
 (n×m)個の前記接続部は、前記脳機能としての正の重み付け機能のみに対応した前記接続部である正専用接続部と、当該脳機能としての負の重み付け機能のみに対応した前記接続部である負専用接続部と、により構成されていることを特徴とする重み符号固定学習装置。
[請求項2]
 請求項1に記載の重み符号固定学習装置において、
 前記正専用接続部と前記負専用接続部とが同数であることを特徴とする重み符号固定学習装置。
[請求項3]
 請求項1又は請求項2に記載の重み符号固定学習装置において、
 前記正専用接続部及び前記負専用接続部が、(n×m)個の前記接続部において一様乱数的に分布していることを特徴とする重み符号固定学習装置。
[請求項4]
 請求項1又は請求項2に記載の重み符号固定学習装置において、
 前記正専用接続部及び前記負専用接続部が、(n×m)個の前記接続部において規則的に分布していることを特徴とする重み符号固定学習装置。
  • Applicant
  • ※All designated countries except for US in the data before July 2012
  • NATIONAL UNIVERSITY CORPORATION HOKKAIDO UNIVERSITY
  • Inventor
  • ASAI Tetsuya
IPC(International Patent Classification)
Please contact us by E-mail or facsimile if you have any interests on this patent.

PAGE TOP

close
close
close
close
close
close