Lizhong Zheng's Globecom 2019 Keynote

  • Share

There are many places that learning algorithms, in particular neural networks, can be used in communication systems, but physical layer problems are often considered harder, or even unsuitable for neural networks. This is mainly because that we traditionally take a model-based approach in communication problems, and often expect strong guarantees on the performance, optimality, robustness, and efficient use of resources, which are all difficult with neural network solutions. The core issue is that unlike image or natural language problems, a communication system is always carefully designed, with structures that are critical in order to achieve the desired performance. However, there is no natural way to use the knowledge about these structures in neural networks, for which part of the design philosophy is to be model agnostic.