Graduate School

Graduate Degrees

I currently attend the University of Texas at Austin North Carolina State University, where I am pursuing a PhD supervised by Prof. Robert W. Heath Jr. My primary research is integrating signal processing and machine learning for wireless communications.

I finished my first year, and I can safely say it was a huge year for me. I took foundational courses in probability, estimation, and machine learning. In spring, I took a course on statistical machine learning, which was both incredibly interesting and quite tough. My research kicked off well, with "Deep Learning-based Carrier Frequency Offset Estimation with 1-Bit ADCs" accepted to SPAWC 2020. I was able to give a recorded presentation of that paper as well, which are two big milestones in my research career. I was also intended to intern at MIT Lincoln Labs, however, that was canceled as a result of COVID-19. Luckily, the wireless team at Facebook was looking for assistance with simulations and AI for developing Open Radio Access Network (ORAN) functionality. My work with Facebook was fruitful and led to continued collaboration and support. We published the first of our work on coverage and capacity optimization (CCO) in ICASSP 2021, which I presented in a recorded format. This paper was especially interesting for considering the Pareto Frontier of CCO as well as comparing Bayesian Optimization and deep reinforcement learning. Later that year, I completed my master's degree at UT and have submitted my first journal paper on multi-sinusoidal parameter estimation from low resolution sampling. We proposed a novel neural architecture that uses successive estimation and cancellation to produce an effecient and powerful estimator that outperforms traditional methods. We also used this work to present a simple learning heuristic: the "learning threshold". This statistcal measure determines what level of generalization or feature learning that a neural network achieves compared to simple distributional learning. In my third (current) year, I have been working on codebook design and become very familiar with 5G NR beam management and feedback. This work is still waiting to be published at VTC 2022 and a future journal, so stay tuned!

Deep Learning-based Carrier Frequency Offset Estimation with 1-Bit ADCs

Ryan M. Dreifuerst, Robert W. Heath Jr, Mandar Kulknari, Jianzhong (Charlie) Zhang

Low resolution architectures are a power efficient solution for high bandwidth communication at millimeter wave and TeraHertz frequencies. In such systems, carrier synchronization is important yet has not received much attention in prior work. In this paper, we develop and analyze deep learning architectures for estimating the carrier frequency of a complex sinusoid in noise from the 1-bit samples of the in-phase and quadrature components. Carrier frequency offset estimation from a sinusoid is used in GSM and is a first step towards developing a more comprehensive solution with other kinds of signals. We train four different deep learning architectures each on five datasets which represent possible training considerations. Specifically, we consider how training with various signal to noise ratios (SNR), quantization, and sequence lengths affects estimation error. Further, we compare each architecture in terms of scalability for MIMO receivers. In simulations, we compare computational complexity, scalability, and mean squared error versus classic signal processing techniques. We conclude that training with quantized data, drawn from signals with SNR between 0-10dB tends to improve deep learning estimator performance across the range of interest. We conclude that convolutional models have the best performance, while also scaling for massive MIMO situations more efficiently than FFT models. Our approach is able to accurately estimate carrier frequencies from 1-bit quantized data with fewer pilots and lower signal to noise ratios (SNRs) than traditional signal processing methods.

Frequency Synchronization for Low Resolution Millimeter-Wave

Ryan M. Dreifuerst, Robert W. Heath Jr, Mandar Kulknari, Jianzhong (Charlie) Zhang

Low resolution data converters can enable power efficient high bandwidth communication at millimeter-wave and terahertz frequencies. Synchronization of such systems is a critical step in accurate decoding, yet current approaches require long block lengths or fail to reach the Cram{\'e}r Rao Bound (CRB). Prior solutions have traditionally been divided into two distinct focuses: algorithms and designed sequences for synchronization. In this paper, we develop a jointly optimized neural architecture for frequency synchronization from configurable sequences and estimators. Our proposed technique uses two neural networks to generate sequences and determine the carrier frequency offset of the sequence after propagating through a channel and applying one-bit quantization. Our simulations show that we can improve estimation performance at low signal to noise ratio (SNR) by up to 8dB at little cost compared to the same estimator without the sequence generator. Our proposed system is fast, efficient, and easily updated, allowing it to handle time-varying systems. In conclusion, we believe further investigation in jointly optimized pilot sequences and estimators will be fundamental to handling signal processing techniques with low resolution data converters.

Optimizing Coverage and Capacity in Cellular Networks using Machine Learning

Ryan M. Dreifuerst et. al

Wireless cellular networks have many parameters that are normally tuned upon deployment and re-tuned as the network changes. Many operational parameters affect reference signal received power (RSRP), reference signal received quality (RSRQ), signal-to-interference-plus-noise-ratio (SINR), and, ultimately, throughput. In this paper, we develop and compare two approaches for maximizing coverage and minimizing interference by jointly optimizing the transmit power and downtilt (elevation tilt) settings across sectors. To evaluate different parameter configurations offline, we construct a realistic simulation model that captures geographic correlations. Using this model, we evaluate two optimization methods: deep deterministic policy gradient (DDPG), a reinforcement learning (RL) algorithm, and multi-objective Bayesian optimization (BO). Our simulations show that both approaches significantly outperform random search and converge to comparable Pareto frontiers, but that BO converges with two orders of magnitude fewer evaluations than DDPG. Our results suggest that data-driven techniques can effectively self-optimize coverage and capacity in cellular networks.

SignalNet: A Low Resolution Sinusoid Decomposition and Estimation Network

Ryan M. Dreifuerst and Robert W. Heath Jr

The detection and estimation of sinusoids is a fundamental signal processing task for many applications related to sensing and communications. While algorithms have been proposed for this setting, quantization is a critical, but often ignored modeling effect. In wireless communications, estimation with low resolution data converters is relevant for reduced power consumption in wideband receivers. Similarly, low resolution sampling in imaging and spectrum sensing allows for efficient data collection. In this work, we propose SignalNet, a neural network architecture that detects the number of sinusoids and estimates their parameters from quantized in-phase and quadrature samples. We incorporate signal reconstruction internally as domain knowledge within the network to enhance learning and surpass traditional algorithms in mean squared error and Chamfer error. We introduce a worst-case learning threshold for comparing the results of our network relative to the underlying data distributions. This threshold provides insight into why neural networks tend to outperform traditional methods and into the learned relationships between the input and output distributions. In simulation, we find that our algorithm is always able to surpass the threshold for three-bit data but often cannot exceed the threshold for one-bit data. We use the learning threshold to explain, in the one-bit case, how our estimators learn to minimize the distributional loss, rather than learn features from the data.