From July 14th to 20th, the 123rd Meeting of MPEG, the ISO/IEC working group for standards for coded representation of digital multimedia data, took place at the Ljubljana Exhibition and Convention Center. Among the emerging topic discussed by international experts is the compressed and interoperable representation of artificial neural networks used in deep learning applications.
Artificial neural networks are used for a wide range of tasks in the areas of multimedia analysis and processing, media coding, data analysis. Their recent success is easy to explain: the ability to process much larger and more complex neural networks than in the past, the availability of the extensive training datasets enable powerful automatic approaches in many areas, including multimedia analysis, surveillance and natural language processing. As trained neural networks contain a large number of parameters, they may become large, which is in particular an issue when deploying to mobile devices or smart cameras. Any application in which a trained neural network is used on a number of devices could thus benefit from a standard for the compressed representation of neural networks.
The MPEG group on compressed neural network representation (NNR) group has thus issued a Call for Evidence to initiate the evaluation of neural network compression technology in three application areas. This group is co-chaired by JOANNEUM RESEARCH Key Researcher Werner Bailer, who also chairs the Austrian working group responsible for contributing to the ISO/IEC standards committee on Artificial Intelligence (SC42). The assessment of these technologies is expected to lead to a standard for the interoperable and compact representation of neural networks.
Links:
The Moving Picture Experts Group