Deep convolutional neural network streaming reasoning method based on time sequence

Abstract

The invention discloses a deep convolutional neural network streaming reasoning method based on a time sequence. The method comprises the following steps: firstly, constructing and initializing a state variable of each layer of network according to parameters of a deep convolutional neural network, and then splicing the state variable and increment input of each layer of network and sending the state variable and increment input into the convolutional neural network layer by layer to calculate and finish reasoning of a single time step; and finally, iteratively updating and recursively finishing the whole streaming reasoning process through state variables and incremental input. According to the method, the characteristic that input features and intermediate calculation variables are partially overlapped in time dimension during deep convolutional neural network streaming reasoning based on a time sequence is utilized, and the state variables are constructed and iteratively updated tostore part of the input features and the intermediate variables which are overlapped in front and back time steps, so that a large amount of repeated operation is reduced; and the computing power andmemory access of the model are effectively reduced under the condition that the performance is not changed.

Description

technical field [0001] The invention belongs to the field of deep convolutional neural networks, and in particular relates to a time series-based deep convolutional neural network streaming reasoning method. Background technique [0002] In recent years, with the rapid development of deep learning, significant progress has been made in related fields such as computer vision, computer hearing, and natural language processing, and deep convolutional neural networks are one of the core technologies in deep learning. However, in the process of continuous breakthroughs in the performance of deep convolutional neural networks, the width, depth, and complexity of the network are also increasing. This also makes the computing overhead and storage overhead of the network increasingly large, making it difficult to complete deployment and real-time operation on embedded and mobile devices such as development boards and smartphones. [0003] For time series represented by speech signal...

AI Technical Summary

Problems solved by technology

In order to ensure the real-time performance of the streaming reasoning, the calculation of each time step needs to be completed within one time step, and the deep convolutional neural network is a neural network with a small network weight but high computational complexity.

Images