Catching Up On 2025-2026 Papers
We’ve been busy! Since last fall, the Living Edge Lab has published seven new papers! So you can catch up on what’s been going on, here’s a quick summary! All of these as well as past publications are available here.
September 2025
y
Wearable Cognitive Assistance (WCA) has the potential to revolutionize daily life with real-time, context-aware guidance, but current performance models often overlook the dynamic nature of human-system interactions, leading to inefficiencies in resource allocation and system responsiveness. In this work, we investigate the implications of the correlated nature of human-system interactions through the development of a novel data-driven methodology for the modeling of task execution times in WCA. We apply this methodology to a WCA application previously shown to exhibit second-order effects between system responsiveness and human performance. Our resulting model presents an improvement in up to 30% with respect to traditional first-order approaches, highlighting the importance of capturing complex behavioral dynamics. These findings raise important questions about the design and optimization of WCA systems and the tools that target them: What are the implications of this correlation for resource allocation and system design in real-world deployments? How can our methodology inform the development of more accurate and adaptive models for WCA applications? By exploring these questions, this research aims to contribute to the development of more efficient and effective WCA systems.
November 2025
We propose a smooth regularization technique that instills a strong temporal inductive bias in video recognition models, particularly benefiting lightweight architectures. Our method encourages smoothness in the intermediate-layer embeddings of consecutive frames by modeling their changes as a Gaussian Random Walk (GRW). This penalizes abrupt representational shifts, thereby promoting low-acceleration solutions that better align with the natural temporal coherence inherent in videos. By leveraging this enforced smoothness, lightweight models can more effectively capture complex temporal dynamics. Applied to such models, our technique yields a 3.8% to 6.4% accuracy improvement on Kinetics-600. Notably, the MoViNets model family trained with our smooth regularization improves the current state of the art by 3.8% to 6.1% within their respective FLOP constraints, while MobileNetV3 and the MoViNets-Stream family achieve gains of 4.9% to 6.4% over prior state-of-the-art models with comparable memory footprints.
The mobility penalty refers to the empirical observation that constraints such as weight, size, energy-efficiency, and thermal dissipation result in mobile devices being much less powerful than static servers of the same technological vintage. This paper examines how real-time AI on mobile devices is negatively affected by the mobility penalty. It characterizes the current state of the art, and proposes an evolutionary approach that side steps the mobility penalty. In this approach, cloudlet-based edge computing serves a catalyst for prototyping and evolving new real-time AI applications for mobile devices without incurring the up-front cost of developing specialized hardware.
December 2025
Despite advances in hardware acceleration, implementing AI on mobile devices is difficult when tight real-time latency bounds have to be met without compromising accuracy. A simple solution is edge offload: using a low-latency wireless network to perform the AI on a nearby cloudlet. This approach also avoids the software engineering effort of downsizing cloud-based AI. In this paper, we experimentally compare on-device and offloaded AI execution by introducing a set of new benchmarks for computer vision tasks. The results show that edge offload is Pareto-optimal across accuracy and latency. It also greatly reduces on-device energy usage.
This paper introduces Black Swan Discovery in Live Learning, a pipelined and iterative semi-supervised machine learning workflow for IoT data collected by static sensors or unmanned robots such as aerial drones, satellites, interplanetary spacecraft, and underwater drones. A black swan is a near out-of-distribution class that is close to some existing classes in feature space, but diverges sufficiently from them to merit a new class label. In the context of extreme class imbalance and severe constraints on data transmission and labeling, this work introduces two approaches to implementing black swan discovery. One approach uses clustering at the edge. The second approach uses active management of the transmission queue. We experimentally compare these approaches, and show that they can both be effective in accelerating black swan discovery.
January 2026
Edge computing supports a wide range of real-time use cases across diverse domains. However, the rapid emergence of Artificial Intelligence (AI) use cases such as smart aging and intelligent video surveillance is introducing complex requirements that existing edge architectures are not designed to handle. Addressing these requirements requires a comprehensive rethink of how edge infrastructure is delivered. This article highlights emerging AI use cases and outlines the architectural and operational requirements and challenges that must be addressed to enable the effective deployment of next-generation AI at the edge.
April 2026
This article examines when a mobile system should view a resource as portable (and hence carried by its user), and when it can be viewed as pervasive (and hence easily found whenever and wherever needed). It shows that the anticipated worst case scenario while mobile is the critical determinant. If the resource is easily and reliably available at all visited locations, then it can be treated as a pervasive resource. If there is reasonable doubt about its availability or attributes at even one visited location, then it has to be treated as portable. Using this resource-centric viewpoint, this article highlights the unique role of the cloud in mobile computing, as well as its limitations. It discusses how some legacy systems can be rearchitected using modern technology, and how emerging real-time mobile AI systems will have to straddle the “carry” versus“find” divide.
More to come in 2026!!