T/AI 115.1-2021 Information Technology, Neural Network Representation and Model Compression, Part 1: Convolutional Neural Networks (CNNs)

Scope

Scope: This document specifies the basic representation methods and compression processes for efficient convolutional neural network models that adapt to various computing requirements. This document is applicable to the research, development, and testing evaluation processes of various convolutional neural network models, as well as their efficient application in the edge and cloud domains. Main technical content: This document aims to provide reference for the representation and compression techniques that may be involved in neural networks, enhancing the reusability of models for users. When using this document, necessary support should be provided for the model's representation methods, and optional support for compression techniques can be provided based on actual application scenarios and technical configurations. Specific support methods will be supplemented by subsequent standards. For the representation methods specified in this standard, no native platform support is required. Support can be provided through conversion or toolkits. The definitions in this document can be transformed into forms and implementations that match specific computing devices and frameworks.

Development Information

Related Standards