Artificial neural networks have been demonstrated for use in a broad range of tasks in multimedia analysis and processing, such as visual and acoustic classification, object and pattern recognition, extraction of multimedia descriptors, or image and video coding. The trained neural networks for these applications contain a large number of parameters (weights), resulting in a considerable amount of data needed to represent the neural network itself. For efficiently transmitting these trained networks to devices (e.g., mobile phones, smart cameras), compression of neural networks is needed. The MPEG standard for compressed representation of neural networks for multimedia content and analysis, currently under Draft International Standard (DIS) ballot, addresses these requirements and provides technologies for parameter reduction and quantization in order to compress entire neural networks.
In emerging application scenarios of neural networks, such as edge-based content processing or federated training, updates of neural networks (e.g., after training on additional data) need to be exchanged. Such updates include changes of the network parameters and may also involve structural changes of the network (e.g., when extending a classification method with a new class). In scenarios like federated training, these updates are more frequent than initial deployments of trained networks, and thus require much more bandwidth over time. However, there is evidence that these updates with respect to a base model can be compressed very efficiently.
MPEG Technical Requirements has thus issued a Call for Proposals (CfP) for compression technology for incremental coding of neural networks, addressing weight and structure updates. The compression technology will be evaluated in terms of compression efficiency and the impact on performance in two use cases. Responses to the CfP will be analyzed at the 134th MPEG meeting. For further information about the call or regarding responses to the call please contact Werner Bailer firstname.lastname@example.org and Dr. Igor Curcio (MPEG Technical Requirements Convenor, email@example.com).