Machine Learning and Optimization for Communications and Deep Networks

In recent years, many remarkable achievements have been made in the field of machine learning. While most of the initial successes were related to image, speech and language recognition, a recent important development has been the application of these techniques to other areas. In particular, communications systems can benefit from applying these techniques. For example, algorithms such as Monte Carlo Markov Chain and Monte Carlo Tree Search have been successfully used in the design of MIMO (i.e., multiple antenna) transceivers. In addition, highly quantized implementations, such as binarized networks, have led to implementations that are well-suited to power-limited mobile platforms. In addition, metaheuristic optimization techniques such the genetic algorithm and others have been used to automatically find highly efficient deep learning architectures, eliminating the need for lengthy and tedious manual experimentation. This lecture will describe these approaches and present some recent design examples. Relationships between the algorithms will be emphasized, and important computational issues will be highlighted. Finally, opportunities for future research in these areas will be suggested.