Computing device and method applied to long short term memory neural network

Abstract

The invention provides a computing device and method applied to a long short term memory neural network. The computing device comprises a computing unit for executing the computation of a forgetting gate function, an input gate function, an output gate function and the state of a memory unit of the long short term memory neural network; and a vector computing unit for obtaining a current output value vector and a current state value vector of the memory unit of the long short term memory neural network on the basis of a previous state value vector of the memory unit of the long short term memory neural network and a current forgetting gate value vector, a current input gate value vector, a current output gate value vector and a current immediate state value vector of the memory unit whichare obtained through computation. By means of the computing method and the computing device which are provided by the invention, the resource utilization rate and the computing efficiency of the longshort term memory neural network can be increased.

Description

technical field [0001] The invention relates to the technical field of deep learning, in particular to a computing device and a computing method applied to a long-short-term memory neural network. Background technique [0002] Long short-term memory neural network (LSTM) is a special recurrent neural network with the ability to learn long-term dependencies, which can be applied to learning language translation, robot control, image analysis, document summarization, speech recognition, handwriting recognition, etc. LSTM can remember information for a long time during the processing of target data. In the process of LSTM application, a processing unit with judgment ability is usually added to the algorithm. The processing unit includes an input gate, a forgetting gate and an output gate. According to the rules, it is judged whether the input information is useful, and the information that conforms to the algorithm certification is retained. The information that does not match...

AI Technical Summary

Problems solved by technology

[0003] However, in the calculation process of the existing LSTM network, since a simple gate circuit structure is usually used for calculation, there is a serial relationship between the multiplication and accumulation process of vectors and weights and the operation of partial gate value vectors, and the utilization rate of computing resources and The rapidity of data processing is difficult to maintain at the same time, and the idle state of computing resources inevitably occurs, which reduces resource utilization. In addition, there are problems in the LSTM network with high memory access times and low operating power consumption.

Images