CAREER: Generalizing Deep Learning for Wireless Communication

This award is funded in whole or in part under the American Rescue Plan Act of 2021 (Public Law 117-2).

Wireless channels in Next Generation (xG) cellular, vehicular (V2X), air-borne, millimeter wave (mmWave) networks are crippled by time-varying impairments that limit their utility in practice. Current art employs spatial multiplexing at the transmitter which is computationally expensive and brittle under statistically non-stationary xG channels, complicating the receiver as well.

At best, these methods are only able to achieve a modest error rate that is inadequate to support high data rate wireless applications like mobile AR/VR/XR, aerial communications and 4K/8K HDR video streaming services. This CAREER research generalizes the architecture of a Deep Learning (DL) based wireless transceiver that will consistently operate with low error rate in all types of wireless channels, but especially outperform the state of the art in future xG channels.

Overall, it is envisioned to achieve 3-5 orders of magnitude improvement in reliability across all types of channels and applications. The education plan focuses on a web-learning platform that augments traditional textbooks with interactive elements, multimedia and adaptive content to promote self-learning. Further, an extended reality platform is envisioned for virtual laboratory experience that currently limits the hands-on aspect of engineering education.

Collectively, the education plan broadens the participation of students beyond the boundaries of the PI’s home institution.

This project expands the understanding and applicability of deep learning (DL) for practical wireless transceivers in four fundamental areas: 1) Reliability: It takes a mathematically principled approach towards understanding the generality of DL models for wireless communications that can adapt to changing wireless environment without compromising reliability; 2) Generality: Current art often lead to over-optimized models that are brittle when exposed to non-stationary changes in the channel state. This research takes a holistic approach by innovating adaptive algorithm for accurate spatio-temporal decomposition of the channel state and pre-condition the waveform for error free communications; 3) Complexity: Low computational complexity of the proposed methods will make DL transceivers easy to reconfigure with minimal to no retraining and operate with guaranteed error performance; and 4) Adaptability: Data dependent, inverse model design and transfer learning will ensure the DL models can adapt quickly to ephemeral channel states without compromising on reliability and complexity.

Finally, the research is made practical by prototype hardware implementation of the transceiver architecture and validated with extensive over-the-air experimentation. Overall, the transmitter and receiver work together to adapt in any and all channel conditions that balances model complexity and error performance.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.