LONG SHORT-TERM MEMORY NEURAL NETWORK CIRCUIT AND CONTROL METHOD

WIPO Patent Application WO/2024/051525

A1

The present application provides a long short-term memory neural network circuit and a control method. The long short-term memory neural network circuit comprises an input circuit, a weight circuit and an activation function circuit, which are electrically connected in sequence, wherein a dynamic memristor in the input circuit superimposes an input at the current moment onto a quantity of state stored at the previous moment, so as to obtain superimposed input quantities at different moments; the weight circuit performs a matrix-vector multiplication and addition operation on a stored weight matrix vector and each output, which is superimposed by the dynamic memristor, of the input circuit, so as to obtain an initial output quantity of a weight matrix; and the activation function circuit performs nonlinear calculation on the initial output quantity of the weight matrix, so as to obtain a voltage output result of the long short-term memory neural network circuit. Therefore, the problem of additional weight and bias configurations being added in order to meet the requirements of a dynamic feedback mechanism, thus resulting in a system loss increase and a time delay, is solved, the usage of weight and bias parameters and caches of a network is reduced, and the frequencies of data writing and reading are reduced, thereby reducing the hardware overheads and time delay of a system.

LI XINYI (CN)
WU HUAQIANG (CN)
QIAN HE (CN)

PCT/CN2023/115473

March 14, 2024

August 29, 2023

Click for automatic bibliography generation  

UNIV TSINGHUA (CN)

G06N3/063

CN115719087A	2023-02-28
CN113988281A	2022-01-28
CN111680792A	2020-09-18
CN108510065A	2018-09-07
US10452472B1	2019-10-22

TSINGYIHUA INTELLECTUAL PROPERTY LLC (CN)

View/Download PDF      PDF Help

Previous Patent: JOINT PREDICTION METHOD AND APPARATUS FOR HYDRAULIC, WIND AND PHOTOVOLTAIC GENERATION POWER

Next Patent: ROCKING CHAIR