Marco Levorato

Accurate stress monitoring in everyday settings requires informative labels without excessive user burden. Ecological Momentary Assessments (EMAs) provide self-reported labels, but poorly timed or frequent prompts can reduce their utility. This paper presents a context-aware active reinforcement learning algorithm for stress detection using photoplethysmography (PPG) from smartwatches and contextual data from smartphones. The algorithm uses the user's current context and response history to decide when to trigger an EMA. We first evaluate context-aware label selection in an offline study and then deploy the approach for real-time EMA triggering in a separate online cohort. Under the reported evaluation, Random Forest F1 increased from 0.21 for the offline cohort to 0.32 for the online cohort; adding contextual features increased it to 0.36. Personalization increased AUC by up to 0.10 across Random Forest, XGBoost, and SVM classifiers. These results support context-aware EMA scheduling and personalized modeling for real-world stress monitoring, while differences between the study cohorts limit causal attribution of the observed performance gains.

Computer vision (CV) tasks are among the most pivotal, yet challenging, operations for uncrewed aerial vehicles (UAVs), especially in mission-critical applications. They require processing complex image data through deep neural networks (DNNs), which demand computational resources far beyond UAVs’ capacity. To address this limitation, Split DNNs offer a promising solution by partitioning the model into: 1) a lightweight Head, deployed on the UAV for rapid, albeit less precise, initial image representations, and 2) a more complex Tail, executed at the network edge for refined, higher accuracy results. However, this solution necessitates transmitting large tensor data from the UAV to the edge server, leading to significant bandwidth consumption. We tackle this challenge by introducing a goal-oriented framework named compressed tensor-based DNN split (CoTeD). Our framework integrates an application- and system-aware optimization model that orchestrates computing and transmission resources in real time. At the UAV, CoTeD dynamically selects relevant tensor information and optimally tradesoff between DNN detection quality and bandwidth consumption, guided by application requirements and system operational conditions. At the edge server, CoTeD reconstructs the tensor, enabling efficient inference by the Tail model. This approach effectively balances bandwidth usage with the quality of the CV task output. Experimental results, obtained through our hardware-software testbed and using datasets with different sizes and characteristics, show that CoTeD can reduce data transmission over the radio link by up to 90% without noticeable loss in object detection quality and inference latency by up to 70% compared to local DNN deployment onboard the UAV. Also, CoTeD yields an inference request success rate of at least 90%, with an increase of 20%–80% compared to direct DNN splitting, static Joint Photographic Experts Group (JPEG) compression, and DNN model quantization.

Developing algorithms for sound classification, detection, and localization requires large amounts of flexible and realistic audio data, especially when leveraging modern machine learning and beamforming techniques. However, most existing acoustic simulators are tailored for indoor environments and are limited to static sound sources, making them unsuitable for scenarios involving moving sources, moving microphones, or long-distance propagation. This paper presents DynamicSound an open-source acoustic simulation framework for generating multichannel audio from one or more sound sources with the possibility to move them continuously in three-dimensional space and recorded by arbitrarily configured microphone arrays. The proposed model explicitly accounts for finite sound propagation delays, Doppler effects, distance-dependent attenuation, air absorption, and first-order reflections from planar surfaces, yielding temporally consistent spatial audio signals. Unlike conventional mono or stereo simulators, the proposed system synthesizes audio for an arbitrary number of virtual microphones, accurately reproducing inter-microphone time delays, level differences, and spectral coloration induced by the environment. Comparative evaluations with existing open-source tools demonstrate that the generated signals preserve high spatial fidelity across varying source positions and acoustic conditions. By enabling the generation of realistic multichannel audio under controlled and repeatable conditions, the proposed open framework provides a flexible and reproducible tool for the development, training, and evaluation of modern spatial audio and sound-source localization algorithms.

Machine learning (ML) can be often distributed, owing to the need to harness more resources and/or to preserve privacy. Accordingly, distributed learning has received significant attention from the literature; however, most works focus on the expected learning quality (e.g., loss) attained and do not consider the distribution thereof. It follows that ML models are not dependable, and may fall short of the required performance in many real-world cases. In this work, we tackle this challenge and propose DepL, a framework attaining dependable learning orchestration: DepL efficiently makes joint, near-optimal decisions concerning (i) which data to use for learning, (ii) the ML models to use – chosen within a set of full-size models and compressed versions thereof – and when to switch from one model to another, and (iii) the clusters of physical nodes to use for the learning. DepL improves over previous works by guaranteeing that the learning quality target (e.g., a minimum loss) is achieved with a target probability, while minimizing the learning (e.g., energy) cost. DepL has provably low polynomial computational complexity and a constant competitive ratio. Further, experimental results using the CIFAR-10 and GTSRB datasets show that it consistently matches the optimum and outperforms state-of-the-art approaches (30% faster learning and 40–80% lower cost).

Next-generation mobile systems will support multiple AI-based applications, each leveraging heterogeneous sensors and data sources through deep neural network (DNN) architectures collaboratively executed within the network. In this context, to minimize the cost of the AI inference task subject to requirements on latency, quality, and – crucially – reliability of the inference process, it is vital to optimize (i) the set of sensors/data sources and (ii) the DNN architecture, (iii) the network nodes executing sections of the DNN, and (iv) the resources to use. To achieve these goals, we leverage dynamic gated neural networks with branches, and propose a novel algorithmic strategy called Quantile-constrained Inference (QIC), based upon quantile-Constrained policy optimization. QIC makes joint, high-quality, swift decisions on all the above aspects of the system, with the aim to minimize inference energy cost. We remark that this is the first contribution connecting gated dynamic DNNs with infrastructure-level decision making. We evaluate QIC using a dynamic gated DNN with stems and branches for optimal sensor fusion and inference, trained on the RADIATE dataset offering Radar, LiDAR, and Camera data, and real-world wireless measurements. Our results confirm that QIC closely matches the optimum and outperforms existing approaches in reducing energy consumption (compute, communication, and total) and application requirements failure by over 70%.

An increasing number of mobile applications leverage deep neural networks (DNN) as an essential component to adapt to the operational context at hand and provide users with an enhanced experience. It is thus of paramount importance that network systems support the execution of DNN inference tasks in an efficient and sustainable way. Matching the diverse resources available at the mobile-edge-cloud network tiers with the applications requirements and the complexity of their, while minimizing energy consumption, is however challenging. A possible approach to the problem consists in exploiting the emerging concept of dynamic DNNs, characterized by multi-branched architectures with early exits enabling sample-based adaptation of the model depth. We leverage this concept and address the problem of deploying portions of DNNs with early exits across the mobile-edge-cloud system and allocating therein the necessary network, computing, and memory resources. We do so by developing a 3-stage graph-modeling method that allows us to represent the characteristics of the system and the applications as well as the possible options for splitting the DNN over the multi-tier network nodes. Our solution, called Feasible Inference Graph (FIN), can determine the DNN split, deployment, and resource allocation that minimizes the inference energy consumption while satisfying the nodes’ constraints and the requirements of multiple, co-existing applications. FIN closely matches the optimum and leads to over 89% energy savings with respect to state-of-the-art alternatives.