Method for constructing incremental LSTM by utilizing training process compression and memory consolidation

Abstract

The invention discloses a method for constructing incremental LSTM by utilizing training process compression and memory consolidation. According to the method, an important input moment in the training process is selected by utilizing the activity of the LSTM gate unit, the training process is compressed to form historical memory, the compressed memory is effectively fused with training on new data, and the network memory is consolidated by utilizing historical information so as to meet the incremental processing requirement on the sequence data. According to the method, the problems of high storage overhead and low training efficiency of an existing LSTM system can be avoided, the training efficiency of the LSTM for dynamically increased sequence data is improved, forgetting of historicalinformation is avoided, and the effectiveness and practicability of the LSTM are guaranteed.

Description

technical field [0001] The invention relates to the field of deep learning of artificial intelligence, in particular to a method for constructing an efficient incremental LSTM by using training process compression and memory consolidation. Background technique [0002] In recent years, with the continuous development of new artificial intelligence technology and the explosive growth of massive data, how to process and analyze these data efficiently, accurately and quickly with the help of new technologies, and tap the huge value contained in them has become a challenging task . Recurrent neural network is a popular deep learning model suitable for the analysis and modeling of sequence data. Long short-term memory network (LSTM) is a variant of cyclic neural network. By adding a gating mechanism to the memory unit, it overcomes the problem of gradient disappearance in traditional cyclic neural network training. It is the most widely used cyclic neural network structure. [...

AI Technical Summary

Problems solved by technology

[0004] In view of the deficiencies in the prior art, this application proposes a method of constructing an incremental LSTM using training process compression and memory consolidation to solve the problem that the existing LSTM cannot effectively perform incremental learning of sequence data and reduce the space for storing data Overhead, while avoiding the forgetting of historical information caused by the learning of new data, improving the practicability of LSTM

Images