User Tools
Table of Contents
Home – Research – Members – Publication – Honor – Courses – Services – News – Openings
NSF/CNS-2003059: MLWiNS: Deep Neural Networks Meet Physical Layer Communications – Learning with Knowledge of Structure
Personnel
- Principal Investigator in the Leading Institution: Lingjia Liu (ECE at VT)
- Principal Investigator in the Collaborative Institution: Lizhong Zheng (EECS at MIT)
- Student Investigator at VT: Jiarui Xu
Abstract
Artificial intelligence (AI) is having a transformational effect in many industries, however, the application of AI and its associated machine learning (ML) tools within wireless networks, while promising, is still in its nascent stages. Some fundamental difficulties remain in applying learning-based data-driven techniques in communication problems: they generally cannot match the results from the more conventional model-based approaches in offering insights to understand entire families of communication problems. In general, both the learning-based data-driven approach and the model-based approach have some advantages, some success stories and some limitations. The objective of this project is to combine these two approaches for communication problems focusing on the physical layer of a wireless network. To be specific, we will identify novel methods to utilize the knowledge of the structure inherent in a wireless network for the learning process to come up new algorithms that can be tailored to specific yet critical applications, and hence have better efficiency and performance guarantees.
Research Thrusts
We choose to explore the use of learning-based techniques, especially deep neural networks (DNNs), in the physical layer problems of a wireless network. This is not because we expect to see some orders of magnitude performance improvement like we often see from high layer problems, but rather since the physical layer has simple models and clear performance benchmarks, which can provide good guidance when we make changes to the learning algorithm to reflect the domain knowledge. Three critical problem domains have been identified to combine domain knowledge with learning algorithms:
- Thrust 1: Symbol detection for MIMO fading interference channel;
- Thrust 2: Massive MIMO with low-resolution ADCs;
- Thrust 3: MIMO-OFDM waveform and non-linear radio frequency effects.
Publication
