0755 - Welcome & Overview of Workshop

0800 - Paper Session #1: Distributed AI/ML at the Resource-Constrained Edge

20 mins - On the Feasibility of Coherent Digital Over-the-Air Computation Based on Empirical Measurements. Alphan Sahin (University of South Carolina, USA)

20 mins - Real-Time UAS Detection Using Neuromorphic Computing of Neuromorphic Camera Events. Brian A Millikan (Southwest Research Institute & Lockheed Martin Corporation, USA); Chau Ton and Howard Yanxon (Southwest Research Institute, USA); Lauren Reinerman-Jones (University of Central Florida & IEMS, USA). 

0840 - Invited Talk: Prof. Timothy Woods, The George Washington University

Title: A Vision for Semantic Resource Elasticity
in Autonomic Cyber-Physical Edge Systems

Bio: Timothy Wood is a professor in the Department of Computer Science at George Washington University. Before joining GW, he received a doctoral degree in computer science from the University of Massachusetts Amherst and a bachelor’s degree in electrical and computer engineering from Rutgers University. His research studies how new virtualization technologies can provide application agnostic tools that improve performance, efficiency, and reliability in cloud computing data centers and software-based networks. His PhD thesis received the UMass CS Outstanding Dissertation Award, his students have voted him CS Professor of the Year, and he has won three best paper awards, a Google Faculty Research Award, and an NSF Career award. 

0910 - Invited Talk: Prof. Mingyue Ji,  University of Florida

Title: Adaptive Discounting of Compression Error for Decentralized Learning

Abstract: Decentralized learning allows distributed nodes to train a shared model through exchanges with their neighbors,   without a central server. In networks with limited bandwidth and sparse connectivity, however, communication compression, slow information propagation, and heterogeneous local data interact to hinder training. This talk presents A-DEFEAT and A-DEFEAT-GT, two algorithms that adapt the correction of compression error using the current update and stored residual norms at each node. Their distinguishing features include an adaptive residual coefficient that requires no discount-parameter search and neighbor model caches reconstructed exactly from the same compressed increments applied locally. A-DEFEAT-GT combines this mechanism with gradient tracking to accommodate arbitrary data heterogeneity without momentum. A-DEFEAT removes the tracker, transmitting one compressed vector per round instead of the two used by BEER and MoTEF (state of the art), with theoretical guarantees under bounded gradient divergence. This approaches improve the dependence on the network spectral gap relative to MoTEF and BEER under stated stability conditions. Experiments across image-classification benchmarks and networks of up to 200 nodes demonstrate strong performance under aggressive compression. Comparisons at equal training rounds and equal communication budgets illustrate both the gains and the tradeoffs between the two variants.

Bio: Dr. Mingyue Ji is an Associate Professor of Electrical and Computer Engineering at the University of Florida and directs the EdgeAI Laboratory. His research focuses on information theory, cloud and edge computing, distributed machine learning, and next-generation wireless communications, networking, and sensing. His honors include the NSF CAREER Award (2022), the IEEE Communications Society Leonard G. Abraham Prize (2019), and several IEEE best paper awards. He serves as an editor for IEEE Transactions on Information Theory and IEEE Communications Magazine and previously served IEEE Transactions on Communications (2020–2025).

0940 - Invited Talk: Prof. Sanghamitra Dutta, University of Maryland 

Title: Miniature Model Replicas: Efficiency via Interpretability

Bio: Sanghamitra Dutta is an assistant professor in the Department of Electrical and Computer Engineering at the University of Maryland College Park. Before joining UMD, she was a senior research associate at JPMorgan Chase AI Research New York in the Explainable AI Centre of Excellence (XAI CoE). She received her Ph.D. and Master's from Carnegie Mellon University and B. Tech. from IIT Kharagpur, all in Electrical and Computer Engineering. Her research interests broadly revolve around efficient and trustworthy machine learning, by bringing in foundational perspectives from information theory, statistics, optimization, and causality. Her research has been published at several machine learning conferences such as NeurIPS, ICML, ICLR, etc., along with leading information-theory venues, featured in New Scientist, Baltimore Sun, and Montreal AI Ethics Brief, and adopted at JPMorgan. She is a recipient of the 2025 NSF MPS SPEED Award, 2024 NSF CAREER Award, 2024 George Corcoran Memorial Award, 2023 Northrop Grumman Seed Grant, 2023 JPMorgan Faculty Award, 2022 Simons Fellowship for Causality, 2021 AG Milnes Outstanding Thesis Award from CMU, and 2019 K&L Gates Presidential Fellowship in Ethics and Computational Technologies. She featured in the 2025 List of 100 Brilliant Women in AI Ethics. 

1010 - Break

1030 - Invited Talk: Prof. I-Hong Hou, Texas A&M University.

Title: Restless Bandits with Individual Penalty Constraints: Near-Optimal Indices and Deep Reinforcement Learning

Abstract: This paper investigates the Restless Multi-Armed Bandit (RMAB) framework under individual penalty constraints to address resource allocation challenges in dynamic wireless networked environments. Unlike conventional RMAB models, our model allows each user (arm) to have distinct and stringent performance constraints, such as energy limits, activation limits, or age of information minimums, enabling the capture of diverse objectives including fairness and efficiency. To find the optimal resource allocation policy, we propose a new Penalty-Optimal Whittle (POW) index policy. The POW index of an user only depends on the user's transition kernel and penalty constraints, and remains invariable to system-wide features such as the number of users present and the amount of resource available. This makes it computationally tractable to calculate the POW indices offline without any need for online adaptation. Moreover, we theoretically prove that the POW index policy is asymptotically optimal while satisfying all individual penalty constraints. We also introduce a deep reinforcement learning algorithm to efficiently learn the POW index on the fly. Simulation results across various applications and system configurations further demonstrate that the POW index policy not only has near-optimal performance but also significantly outperforms other existing policies.

Bio: I-Hong Hou is a Professor in the ECE Department of the Texas A&M University. He received his Ph.D. from the Computer Science Department of the University of Illinois at Urbana-Champaign. His research interests include wireless networks, edge/cloud computing, and reinforcement learning. His work has received the 2025 IEEE Communications Society William R. Bennett Prize, the Best Paper Award from ACM MobiHoc 2017 and ACM MobiHoc 2020, and Best Student Paper Award from WiOpt 2017. He has also received the C.W. Gear Outstanding Graduate Student Award from the University of Illinois at Urbana-Champaign, and the Silver Prize in the Asian Pacific Mathematics Olympiad.

1100 - Invited Talk: Prof. Emrah Akyol,  SUNY Binghamton. 

Title: Strategic Communication and Coordination in Resource-Constrained Multi-Agent Systems

Abstract: Distributed intelligent systems must make decisions with limited communication, imperfect observations, and agents whose objectives may not coincide. In such settings, communication is not simply about reproducing the source data accurately but also the decisions it induces. This talk brings together information theory and game theory to examine three aspects of this problem. First, I will discuss strategic rate-distortion formulations that characterize how a sender allocates a limited rate budget while anticipating a receiver’s response. For Gaussian models, these problems admit tractable posterior-covariance characterizations. Second, I will show how privacy constraints reshape disclosure and can, under suitable conditions, improve the receiver’s performance by redirecting information toward its task. Finally, I will examine distributed coordination through log-linear learning over noisy communication networks, highlighting how channel reliability and the choice between individual noisy observations and averaged information affect long-run coordination. These three results illustrate the importance of considering communication constraints, incentives, and local decision rules jointly when designing resource-constrained multi-agent systems. 

Bio: Emrah Akyol is an Associate Professor of Electrical and Computer Engineering at Binghamton University, State University of New York. He received his Ph.D. in Electrical and Computer Engineering from the University of California, Santa Barbara, in 2011. Before joining Binghamton in 2017, he held postdoctoral positions at the University of Southern California and the University of Illinois Urbana-Champaign. His research lies at the intersection of information theory, game theory, and control, with an emphasis on strategic communication, privacy, and decision-making in networked systems. He received an NSF CAREER Award in 2021 and is a Senior Member of the IEEE. 

1130 - Invited Talk: Prof. Rafael D'Oliveira, Clemson University 

Title: Secure Single-Server Matrix Multiplication

Abstract: Matrix multiplication is at the core of many machine learning workloads, and resource-constrained devices often cannot afford to compute it locally. Outsourcing it to a more powerful server solves the computation problem, but it raises a new one: the security of the data. Secure distributed matrix multiplication addresses this with information-theoretic guarantees, but only when the number of colluding servers is bounded. When all servers collude, they are effectively a single server, and the question becomes what information-theoretic security we can still provide while keeping the user's work well below that of computing the product locally. In this talk, we explore this single-server setting.

Bio: Rafael G. L. D'Oliveira is an Assistant Professor at the School of Mathematical and Statistical Sciences at Clemson University. He received a B.S. and an M.S. degree in mathematics and a Ph.D. degree in applied mathematics from the University of Campinas in Brazil in 2009, 2012, and 2017. He was a postdoctoral research associate with the Research Laboratory of Electronics at the Massachusetts Institute of Technology from 2020 to 2022, with Rutgers University from 2018 to 2019, and with the Illinois Institute of Technology in 2017. He did a research internship at Telecom Paristech from 2015 to 2016. His research interests include Privacy and Security, Distributed Computing, Coding Theory, and Information Theory.

1200 - Closing Workshop Remarks & Workshop Adjourns