Project Abstract

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.
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.
5. 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.

Intellectual Merit

In recent years, Deep learning (DL) has been applied to wireless communication systems, but in silos of innovation that has limited its utility across a wide variety of applications and channels. The proposed CAREER research is the first to view DL transceivers as a generic waveform processing hardware that not only outperform existing designs but is also able to adapt in extremely dynamic and non-stationary wireless environments. Central to realizing this, is a novel, low complexity DL based spatio-temporal channel decomposition apparatus that pre-compensates the transmission to remove hidden correlations in the channel by an adaptive analysis of the channel state information. This vastly simplifies the receiver architecture, which now operates within chosen bounds of performance while being malleable to the changes in the underlying wireless data samples. Collectively, the proposal will make real-time, data-driven DL ubiquitous and universally applicable in wireless communications that is firmly rooted in provable mathematical principles, something that is currently missing in the literature

Research Objectives

In order to realize the research vision we need to leverage the computational advantage of DL to iterativelyreinforce the knowledge of hidden correlations in the channel and pre-compensate the waveform at the transmitter. This allows for a generic neural network based receiver to operate with a deterministic error bound for any and all wireless environment. Therefore, the objective of this proposal are:

1. To extract meaningful characterizations of multi-modal channels at the transmitter that uniquely define the future stochastic behavior of the channel, that is also computationally efficient.
2. To pre-compensate the waveform by employing the latent representation of the wireless channel to guarantee the error performance at the receiver for all types of channel.
3. To establish a mathematical and functional equivalence between a classical and generalized DL receiver and generate optimal structures with bounded error performance for different wireless channels.
4. To learn and predict the future channel states by employing a data-driven and policy based learning to adapt to NS channels with unknown distributions

Education Objectives

The educational objectives of the CAREER proposal are shaped by the seismic shift in higher education by COVID–19 pandemic that has forced all academicians to re-calibrate their learning and teaching strategies:

1. To expand the understanding of communication systems theory by creating a knowledge bridge to deep learning with focus on explainable data-driven signal processing.
2. To innovate teaching modalities for personalized and interactive learning for classrooms of the future, including piloting instruction in extended reality (XR) environments.
3. To prepare students for life-long learning, create opportunities for traditionally under-represented groups and promote an inclusive learning environment within my research lab as well as in the department.

Publications and Outcomes

1. Z. Zou, M. Careem, A. Dutta and N. Thawdar, “Joint Spatio-Temporal Precoding for Practical Non-Stationary Wireless channels,” in IEEE Transactions on Communications, vol. 71, no. 4, pp. 2396-2409, April 2023. [pdf] .
2. Z. Zou and A. Dutta, "Multi-dimensional Eigenwave Multiplexing (MEM): A General Modulation Beyond OTFS (Under Review).

1. M. Careem, A. Dutta and N. Thawdar, "On Equivalence of Neural Network Receivers," IEEE International Conference on Communications (ICC), Montreal, 2021. [pdf]
2. Z. Zou, M. Careem, A. Dutta and N. Thawdar, "Unified Characterization and Precoding for Non-Stationary Channels," IEEE International Conference on Communications (ICC), Seoul, 2022. (Best Paper Award), [pdf]
3. Z. Zou and A. Dutta, "Mutidimensional Eigenwaves Multiplexing Modulation for Non-stationary Channels," IEEE Global Communications Conference (GLOBECOM), Kuala Lumpur, 2023.[pdf]
4. Z. Zou and A. Dutta, "Capacity Achieving by Diagonal Permutation for MU-MIMO channels," GLOBECOM - IEEE Global Communications Conference (GLOBECOM), Kuala Lumpur, 2023. [pdf]
5. Z. Zou, I. Amarasekara and A. Dutta, "Learning to Decompose Asymmetric Channel Kernels for Generalized Eigenwave Multiplexing," IEEE Conference on Computer Communications (INFOCOM), Vancouver, 2024. [pdf]
6. I. Amarasekara, Z. Zou and A. Dutta, "Adaptive Neural Network for Eigen-Decomposition of Multi-dimensional Channel Kernels," IEEE 99th Vehicular Technology Conference (VTC-Spring), Singapore, 2024,[pdf]

1. M. A. Abdul Careem, “Architecting Future Multi-Modal Networks: Coexistence, Generalization & Testbeds”, Ph.D. Dissertation (May 2023), University at Albany, SUNY. [pdf]
2. Z. Zou, “Waveforms for xG Non-stationary Channels” Ph.D. Dissertation (expected 2025), University at Albany, SUNY.

Technologies or Techniques:
High Order Generalized Mercer’s theorem (HOGMT) is proposed in our work [ICC-22], which is derived from Mercer’s theorem and allows for the decomposition of multi-dimensional processes into eigenfunctions. Significantly, it is the first decomposition method applicable to non-stationary channels. Building on this, we have further proposed characterization, precoding, and modulation methods for non-stationary channels and the results have been disseminated through the publications mentioned in this report. These innovative approaches addressed “How to decompose a non-stationary channel and what is the corresponding signaling scheme”, which has been acknowledged as an open problem in the literature and also mentioned as one of the goals of this project.