Standard: MPEG-7
Part: 17

To define tools for compression of neural networks for multimedia applications and representing the resulting bitstreams for efficient transport.

Editions

Edition - 1: Compression of neural networks

Publication Year: 0
Status: published
Motivations: Recently (deep) neural networks (NNs) have become a widely applied method in many application areas, including signal processing and multimedia. Classification methods, feature extractors and encoding methods based on NNs often outperform hand-crafted approaches. In many applications the trained NNs (which may contain large amounts of data) need to be transmitted to other systems or terminal devices (with possibly limited computing capabilities), where they are used for inference and/or are updated with local data. Thus efficient representations for exchanging NNs are required.
Objectives: To study existing representations of NNs, the state of the art of NN compression methods, and the processing flows of training and deploying NNs to a range of (generic or dedicated) hardware platforms, to identify interfaces where a standard compressed NN representation is needed and the define the requirements for such a representation.

Edition - 2: Compressed representation of neural networks for multimedia content description and analysis

Publication Year: 0
Status: published
Motivations:
Objectives:

Edition - 3: Compression of neural networks

Publication Year: 0
Status: ongoing
Motivations: Recently (deep) neural networks (NNs) have become a widely applied method in many application areas, including signal processing and multimedia. Classification methods, feature extractors and encoding methods based on NNs often outperform hand-crafted approaches. In many applications the trained NNs (which may contain large amounts of data) need to be transmitted to other systems or terminal devices (with possibly limited computing capabilities), where they are used for inference and/or are updated with local data. Thus, efficient representations for exchanging NNs are required.
Objectives: To study existing representations of NNs, the state of the art of NN compression methods, and the processing flows of training and deploying NNs to a range of (generic or dedicated) hardware platforms, to identify interfaces where a standard compressed NN representation is needed and the define the requirements for such a representation.

Whitepapers

Publication date	Title
2022-01-25	White Paper on Neural Network Coding
2024-01-29	White paper on Neural Network Compression

Meeting documents

MPEG 145

Publication date	Title
2024-01-28	Application and Verification of NNC in Different Use Cases
2024-01-29	White paper on Neural Network Compression

MPEG 143

Publication date	Title
2023-07-22	Application and Verification of NNC in Different Use Cases

MPEG 137

Publication date	Title
2022-01-25	White Paper on Neural Network Coding

MPEG 133

Publication date	Title
2021-02-01	Clarifications about NNR evaluation framework

MPEG 132

Publication date	Title
2020-10-23	Evaluation Framework for Compression of neural networks for multimedia content description and analysis
2020-10-23	Call for Proposals on Incremental Compression of Neural Networks for multimedia content description and analysis

MPEG 130

Publication date	Title
2020-05-19	Working Draft 4 of Compression of Neural Networks for Multimedia Content Description and Analysis
2020-04-25	Call for Incremental NNR Test Materials

MPEG 129

Publication date	Title
2020-04-19	Call for Incremental NNR Test Materials

MPEG 127

Publication date	Title
2019-07-28	Evaluation Framework of Compression of Neural Networks for Multimedia Content Description and Analysis
2019-07-13	Use cases and requirements for neural network compression for media content description and analysis