Meeting the latency requirements of 5G networks requires massive parallelization. We'll discuss how to parallelize and map certain radio access network (RAN) functions to GPU architectures to achieve orders-of-magnitude acceleration. We'll describe how to realize selected RAN functions using online machine learning methods. We'll also explore the possibility of a machine learning function orchestrator (MLFO) in the context of end-to-end network slicing where deep neural networks are an interesting option. Our talk will use findings of the ITU-T focus group on machine learning for 5G to explore the challenge of implementing MLFO, leading to new mobile network architectures.

In current wireless networks, most algorithms are iterative and might not be able to meet the requirements of some 5G technologies such as ultra-reliable low-latency communication within a very low latency budget. For instance, requiring and end-to-end latency below 1ms, many signal processing tasks must be completed within microseconds. Therefore, only a strictly limited number of iterations can be performed, which may lead to uncontrollable excessive errors. We argue in favor of formulating the underlying optimization problems as convex feasibility problems in order to enable massively parallel processing on GPUs for online learning for fast and robust tracking. Moreover, convex feasibility solvers allow for an efficient incorporation of context information and expert knowledge, and can provide robust results based on relatively small data sets. Our approach has numerous applications, including channel estimation, peak-to-average power ratio (PAPR) reduction in Orthogonal Frequency Division Multiplexing (OFDM) systems, radio map reconstruction, beam forming, localization, and interference reduction. We show that they can greatly benefit from the parallel architecture of GPUs.