Lightning-fast modulation classification with hardware-efficient neural networks
The ever-growing demand for data is driving a need for improved radio efficiency for 5G and beyond. Automatic Modulation Classification (AMC), which monitors the RF spectrum for different modulation schemes, is a key part of this. Prior works have successfully applied deep learning to AMC, demonstrating competitive recognition accuracy for a variety of modulations and SNR regimes using deep neural networks (DNNs). To unleash the full potential of this approach in real-world applications, the next step is to deploy this technology with ultra-low latency and high throughput. This requires specialized, hardware-efficient DNNs that take both recognition accuracy and the computational cost into account.
In this talk, we introduce the challenge “Hardware-Efficient Modulation Classification on RadioML”, where the goal is to explore hardware-efficient DNNs on the RadioML 2018 dataset. For this challenge, we provide a repository with a DNN training example, and a quantization-aware training library in PyTorch with a built-in hardware cost metric correlated with achievable throughput and latency. We encourage participants to explore different topologies, quantization to few bits in both weights and activations and pruning in order to create hardware-efficient DNNs that can scale performance to unprecedented levels. Winners will be selected on the basis of the achieved hardware cost that achieves recognition accuracy over a certain threshold.