Method and apparatus for performing operations in convolutional neural networks

Abstract

The present application relates to a method and apparatus for performing operations in a convolutional neural network, the method comprising: splitting weight parameters of selected layers in a convolutional neural network to obtain an array of operational parameters comprising a plurality of operational parameters; using Each operation parameter in the obtained array of operation parameters performs an operation of the selected layer to obtain an array of partial operation results including a plurality of partial operation results; and based on the obtained array of partial operation results, generating one or more of the selected layers. Multiple output data. Through this method, the execution efficiency of operations in the convolutional neural network can be improved.

Description

technical field [0001] The present application generally relates to the technical field of artificial convolutional neural networks, and in particular to methods and apparatus for performing operations in convolutional neural networks. Background technique [0002] Deep learning technology based on convolutional neural network has been widely used in different fields such as image recognition, video analysis, natural language processing, and assisted driving. [0003] A convolutional neural network can contain multiple layers. In each layer, the convolution operation of the layer is performed on the input data (also called input feature data) of the layer using the weight parameters of the layer to obtain the corresponding output data (also called activation value or output feature data). [0004] In a convolutional neural network, each feature data can have a certain width and height, and can have one or more channels, and each channel can carry a type of information abou...

AI Technical Summary

Problems solved by technology

However, as the forward inference process of the convolutional neural network proceeds, the size of the weight parameters of each layer may become larger and larger, for example, may have a larger number of channels and / or a larger number of cores

The weight parameters at a certain layer are too large to be fully cached in high-speed memory (e.g., cache memory within or associated with the processor) associated with the processor used to perform the operations in the neural network , which prevents the correct and / or efficient execution of the layer's operations

Images