July 24, 2019

Fraunhofer HHI successfully contributes compression method for Neural Networks to MPEG standardization

One of the key research topics at the Fraunhofer Heinrich-Hertz-Institute HHI is the efficient compression of Neural Networks. At the last standardization meeting of the Moving Picture Experts Group (MPEG), the compression method DeepCABAC (context-adaptive binary arithmetic coding for deep neural network compression), developed by researchers of the institute, was selected as the main basic model for standardizing the compression and representation of Neural Networks.

Neuronal Networks are already shaping daily life because they are the basis for many current approaches to Artificial Intelligence (AI). AI systems are an essential component of future-oriented technologies such as autonomous driving and also play an important role in the next-generation mobile communications standard 5G. However, due to their high number of parameters (weights), they still require a lot of computing capacity and RAM, so that they are difficult to integrate for example into mobile phones without strong compression.

To solve this problem, researchers at Fraunhofer HHI have developed a compression method for Neural Networks: DeepCABAC. Researchers from the two fields of Video Coding and Machine Learning cooperate closely to develop DeepCABAC. By exchanging knowledge and experience, the many years of expertise in the field of video coding could also be used for the compression of Neural Networks. In this way, lossless compression of the sparsified and quantized weights has been built on the method of CABAC, which was developed by Fraunhofer HHI. CABAC is part of the worldwide dominating video coding standards H.264/AVC and H.265/HEVC. Today every second bit on the Internet is generated on the basis of CABAC technology.

Through its adaptive, context-based rate modeling, DeepCABAC allows an optimal quantization and coding of the weight matrices of the Neural Network and thus a very strong compression without performance losses. In many cases, compression of less than five percent of the original size of the Neuronal Network is achieved without significant impairment of the predictive quality ("intelligence") of the network.

At the penultimate MPEG meeting in March 2019, nine proposals for the development of an international standard for the compression of Neural Networks were submitted, including the DeepCABAC proposal of the Fraunhofer HHI. After extensive core experiments and evaluations of all submitted proposals, a first test model for the future standard was defined at the last MPEG meeting. In the process, DeepCABAC was selected as the basic compression technology for two of the three key technological components in the neural network compression pipeline. The Fraunhofer HHI proposal is thus spearheading the further development of the new compression standard. The final draft of the international standard is scheduled for submission in April 2021.