Deep ensembles (DE) have emerged as a powerful approach for quantifying predictive uncertainty and distinguishing its aleatoric and epistemic components, thereby enhancing model robustness and reliability. However, their high computational and memory costs during inference pose significant challenges for wide practical deployment. To overcome this issue, we propose credal ensemble distillation (CED), a novel framework that compresses a DE into a single model, CREDIT, for classification tasks. Instead of a single softmax probability distribution, CREDIT predicts class-wise probability intervals that define a credal set, a convex set of probability distributions, for uncertainty quantification. Empirical results on out-of-distribution detection benchmarks demonstrate that CED achieves superior or comparable uncertainty estimation compared to several existing baselines, while substantially reducing inference overhead compared to DE.

Effectively estimating the uncertainty attached to neural network predictions thus becomes essential to improve robustness, reliability, and trustworthiness. This paper provides an overview of various methodologies for representing, quantifying, and distinguishing two major types of uncertainties (namely, 'aleatoric' and 'epistemic' uncertainty) in neural networks. The review covers classical probabilistic techniques such as Bayesian neural networks and deep ensembles, methods from generalized probability that leverage uncertainty representations such as Dirichlet distributions, belief functions, random sets, probability intervals, and credal sets, among others. Additionally, interval-based approaches employing interval models are also examined. We discuss the strengths and limitations of various methodologies and identify promising research directions for potential future exploration.

This paper presents an innovative approach, called credal wrapper, to formulating a credal set representation of model averaging for Bayesian neural networks (BNNs) and deep ensembles (DEs), capable of improving uncertainty estimation in classification tasks. Given a finite collection of single predictive distributions derived from BNNs or DEs, the proposed credal wrapper approach extracts an upper and a lower probability bound per class, acknowledging the epistemic uncertainty due to the availability of a limited amount of distributions. Such probability intervals over classes can be mapped on a convex set of probabilities (a credal set) from which, in turn, a unique prediction can be obtained using a transformation called intersection probability transformation. In this article, we conduct extensive experiments on several out-of-distribution (OOD) detection benchmarks, encompassing various dataset pairs (CIFAR10/100 vs SVHN/Tiny-ImageNet, CIFAR10 vs CIFAR10-C, CIFAR100 vs CIFAR100-C and ImageNet vs ImageNet-O) and using different network architectures (such as VGG16, ResNet-18/50, EfficientNet B2, and ViT Base). Compared to the BNN and DE baselines, the proposed credal wrapper method exhibits superior performance in uncertainty estimation and achieves a lower expected calibration error on corrupted data.

Effective uncertainty estimation is becoming increasingly attractive for enhancing the reliability of neural networks. This work presents a novel approach, termed Credal-Set Interval Neural Networks (CreINNs), for classification. CreINNs retain the fundamental structure of traditional Interval Neural Networks, capturing weight uncertainty through deterministic intervals. CreINNs are designed to predict an upper and a lower probability bound for each class, rather than a single probability value. The probability intervals can define a credal set, facilitating estimating different types of uncertainties associated with predictions. Experiments on standard multiclass and binary classification tasks demonstrate that the proposed CreINNs can achieve superior or comparable quality of uncertainty estimation compared to variational Bayesian Neural Networks (BNNs) and Deep Ensembles. Furthermore, CreINNs significantly reduce the computational complexity of variational BNNs during inference. Moreover, the effective uncertainty quantification of CreINNs is also verified when the input data are intervals.

Interactive technology and games provide promising methods for wheelchair skills training. This paper introduces RollAbility, a rehabilitative game for powered wheelchair skills training aimed at children and teenagers with complex movement disorders. The game combines clinical best practices and standardized training protocols with player experience goals. Developed through an iterative process with rehabilitation experts and game designers, RollAbility utilizes the Wheelchair Skills Training Program [Kirby et al. '23] and integrates key insights from Aufheimer's [CHI '23] motivation studies in physical therapy to ensure both therapeutic and engaging gameplay. Evaluated through exploratory sessions with children, therapists, and clinical experts, results show that RollAbility effectively merges clinical protocols with an engaging game format. However, balancing player autonomy and therapist guidance remains a critical consideration. This work offers a blueprint for designing therapeutic games that align clinical protocols with engaging player experiences.

This paper introduces an innovative approach to classification called Credal Deep Ensembles (CreDEs), namely, ensembles of novel Credal-Set Neural Networks (CreNets). CreNets are trained to predict a lower and an upper probability bound for each class, which, in turn, determine a convex set of probabilities (credal set) on the class set. The training employs a loss inspired by distributionally robust optimization which simulates the potential divergence of the test distribution from the training distribution, in such a way that the width of the predicted probability interval reflects the epistemic uncertainty about the future data distribution. Ensembles can be constructed by training multiple CreNets, each associated with a different random seed, and averaging the outputted intervals. Extensive experiments are conducted on various out-of-distributions (OOD) detection benchmarks (CIFAR10/100 vs SVHN/Tiny-ImageNet, CIFAR10 vs CIFAR10-C, ImageNet vs ImageNet-O) and using different network architectures (ResNet50, VGG16, and ViT Base). Compared to Deep Ensemble baselines, CreDEs demonstrate higher test accuracy, lower expected calibration error, and significantly improved epistemic uncertainty estimation.

Unhealthy dietary habits are considered as the primary cause of various chronic diseases, including obesity and diabetes. The automatic food intake monitoring system has the potential to improve the quality of life (QoL) of people with diet-related diseases through dietary assessment. In this work, we propose a novel contactless radar-based approach for food intake monitoring. Specifically, a Frequency Modulated Continuous Wave (FMCW) radar sensor is employed to recognize fine-grained eating and drinking gestures. The fine-grained eating/drinking gesture contains a series of movements from raising the hand to the mouth until putting away the hand from the mouth. A 3D temporal convolutional network with self-attention (3D-TCN-Att) is developed to detect and segment eating and drinking gestures in meal sessions by processing the Range-Doppler Cube (RD Cube). Unlike previous radar-based research, this work collects data in continuous meal sessions (more realistic scenarios). We create a public dataset comprising 70 meal sessions (4,132 eating gestures and 893 drinking gestures) from 70 participants with a total duration of 1,155 minutes. Four eating styles (fork & knife, chopsticks, spoon, hand) are included in this dataset. To validate the performance of the proposed approach, seven-fold cross-validation method is applied. The 3D-TCN-Att model achieves a segmental F1-score of 0.896 and 0.868 for eating and drinking gestures, respectively. The results of the proposed approach indicate the feasibility of using radar for fine-grained eating and drinking gesture detection and segmentation in meal sessions.

Eating speed is an important indicator that has been widely investigated in nutritional studies. The relationship between eating speed and several intake-related problems such as obesity, diabetes, and oral health has received increased attention from researchers. However, existing studies mainly use self-reported questionnaires to obtain participants' eating speed, where they choose options from slow, medium, and fast. Such a non-quantitative method is highly subjective and coarse at the individual level. This study integrates two classical tasks in automated food intake monitoring domain: bite detection and eating episode detection, to advance eating speed measurement in near-free-living environments automatically and objectively. Specifically, a temporal convolutional network combined with a multi-head attention module (TCN-MHA) is developed to detect bites (including eating and drinking gestures) from IMU data. The predicted bite sequences are then clustered into eating episodes. Eating speed is calculated by using the time taken to finish the eating episode to divide the number of bites. To validate the proposed approach on eating speed measurement, a 7-fold cross validation is applied to the self-collected fine-annotated full-day-I (FD-I) dataset, and a holdout experiment is conducted on the full-day-II (FD-II) dataset. The two datasets are collected from 61 participants with a total duration of 513 h, which are publicly available. Experimental results show that the proposed approach achieves a mean absolute percentage error (MAPE) of 0.110 and 0.146 in the FD-I and FD-II datasets, respectively, showcasing the feasibility of automated eating speed measurement in near-free-living environments.

Wheelchair skills training is key for achieving greater personal mobility, and interactive solutions offer the potential for a more engaging training experience. In this paper, we examine the tensions between therapeutic requirements and player engagement, through the prototypical development of a wheelchair skills training game for children with Cerebral Palsy. Using an iterative and interdisciplinary design process, exercises from the Wheelchair Skills Program manual are implemented as interactive missions in an open-world, medieval-themed game. The game contains a custom input system that directly uses the player’s wheelchair input peripherals. In bridging the gap between clinical requirements and player engagement, we discuss the challenge of repetition and feedback in therapy, and implications for future work in game design.

Food intake monitoring plays an important role in personal dietary systems. Numerous approaches have been proposed to automatically detect eating gestures using various sensors and machine learning. However, existing eating gesture detection approaches mainly focus on meal sessions. Such a task is still challenging in free-living environments due to longer monitoring duration and more non-feeding activities. This paper proposes a wearable Inertial Measurement Unit (IMU) based method to detect eating and drinking gestures in free-living environments. Two important factors that impede intake gesture detection in free-living environments are addressed: 1) how to handle IMU data from two hands, and 2) what is the impact of downsampling sensor data on performance. To integrate two-hand data, we propose a solution that combines hand mirroring and temporal concatenation techniques. The multi-stage temporal convolutional network (MS-TCN) is applied to effectively recognise intake gestures. A dataset contains 12 subjects with 67.5 h data is collected for validation. Moreover, IMU data with different sampling frequencies are processed to test performance. Validated by Leave-One-Subject-Out (LOSO) method, our approach (with 16 Hz sampling frequency) achieves a segmental F1-score of 0.826 and 0.893 for recognizing eating and drinking gestures, respectively. Results show that the proposed solution outperforms existing two-hand data combination approaches. Moreover, in our case, a higher sampling frequency does not always mean better performance.

While it is widely assumed that games can engage patients in therapy through their inherent ‘motivational pull’, relatively little attention has been paid to what HCI games research can learn from strategies employed by therapists. We address this gap by leveraging Self-Determination Theory (SDT) and its mini-theories Basic Psychological Needs Theory and Organismic Integration Theory as a theoretical lens on physical therapy for children and adolescents. Results from in-depth interviews with twelve therapists show that they carefully adjust sessions to allow patients to experience competence, making more comprehensive adjustments than currently offered by games. Additionally, we highlight how therapists leverage their relationship with patients to support motivation, but struggle to reconcile meaningful experiences of autonomy with therapeutic goals. On this basis, we reflect on implications for researchers and designers who create games for physical therapy, and the potential of SDT to provide a foundation for game design and therapeutic practice.

While the current machine fault diagnosis is affected by the rarity of cross conditional fault data in practice, efficient implementation of these diagnosis models on resource constrained devices is another active challenge. Given such constraints, an ideal fault diagnosis model should not be either generalizable across the shifting domains or lightweight, but rather a combination of both, generalizable while being minimalistic. Preferably being uninformed about the domain shift. Addressing these computational and data centric challenges, we propose a novel methodology, Convolutional Auto-encoder and Nearest Neighbors based self adaptation (SCAE-NN), that adapts its fault diagnosis model to the changing conditions of a machine. We implemented SCAE-NN for various cross-domain fault diagnosis tasks and compared its performance against the state-of-the-art domain invariant models. Compared to the SOTA, SCAE-NN is at least 6− 7% better at predicting fault classes across conditions, while being more than 10 times smaller in size and latency. Moreover, SCAE-NN does not need any labelled target domain data for the adaptation, making it suitable for practical data scarce scenarios.

Bluetooth Low Energy is a wireless communication protocol widely used in Internet of Things systems. As a popular protocol operating in the 2.4 GHz frequency band, it is constantly confronted with interference challenges. For instance, in a Bluetooth Low Energy network, all connections see each other as a source of interference. In this paper, the interference and reliability issues are investigated among Bluetooth Low Energy connections. The impact of various factors on the interference, such as connection parameters and network size, is shown through simulation. The trade-off between application level throughput, i.e. goodput, and reliability in a Bluetooth Low Energy network is described, which can be used in the design or deployment of Bluetooth Low Energy devices or networks.

Bluetooth Low Energy is one of the most popular wireless protocols in the 2.4 GHz frequency band nowadays. There have been multiple studies investigating its performance under interference. However, these studies mostly follow some commonly adopted assumptions. For instance, the location of the interference source does not vary the performance metrics of the connection; the data exchange scheme does not impact the performance metrics of the connection; the performance metrics are the same on both sides of a Bluetooth Low Energy connection. Unfortunately these assumptions are never validated. As a result, this paper aims to verify/challenge these commonly adopted assumptions through experiments. According to the experiment results, the impact of the device role in the BLE connection, the data exchange scheme within the BLE connection, and the location of the interference source is investigated and discussed. This paper should be considered as a cornerstone for other research related to Bluetooth Low Energy performance under interference. It can also be used as a preliminary guideline when deploying Bluetooth Low Energy communication in an interference environment.

We noted that communication performance of a Bluetooth Low Energy (BLE) connection is heavily affected by the radio interference from other connections or networks. Reliability is becoming a key requirement in BLE for its use in various Internet of Things applications. Hence, there is a widely recognized need for an in-depth study to reveal the parameters impacting BLE reliability under such radio interference. In this article, we investigate how transmission parameters, e.g., number of packets and packet transmission time, influence reliability of the BLE protocol. Specifically, a mathematical model is presented to explore the impact of the transmission parameters on the reliability of a BLE pair under interference caused by other pairs. This mathematical model is able to show the reliability issues from both the side of the BLE connection under interference and the side of the interference itself. The model is validated and novel insights on the common usage of BLE parameters by a wide range of experimental evaluations are provided. Experimental results highlight the correctness of the mathematical model, thus quantify the interplay between transmission parameters and coexistence, also the influence of other related parameters. This research provides a design-level or system-level insight in BLE usage and deployment.

With the emergence of mobile devices having enough resources to execute real-time ML inference, deployment opportunities arise on mobile devices while keeping privacy-sensitive data close to the source and reducing server load. Moreover, offloading inference to a cloud server facilitates deployment of neural network-based applications on resource-constrained devices. Depending on the application goals and execution context of the application, the optimal deployment on either cloud server or mobile device varies during the lifetime of an application. In this paper, we propose a context-aware middleware that enables optimization of deployed application software to satisfy the application’s functional goals in accordance with changing execution context and environmental conditions. We facilitate system design through the abstraction of deployed software components as states and make use of finite state machines and contextual triggers to model the reconfiguration of the system. We evaluate our framework using a real-world nutritional monitoring application via food image recognition deployed in a two-tier mobile and cloud architecture. We compare the proposed solution with various static deployments of the application and show that our approach can react to changing application goals at run-time in order to reduce server load and thereby increase scalability.

This paper starts from the observation that mobile and edge devices are powerful enough to execute Machine Learning (ML) application components, which in turn creates opportunities to keep privacy-sensitive data close to its source. Composing and deploying a distributed ML application is far from trivial because the optimal configuration depends on the application’s goals and execution context, both of which may change throughout its lifetime. Prior research on context-aware reconfigurations in ML based applications offer limited capabilities for dynamically migrating software components between mobile, edge and cloud devices. In this paper, we propose a context-aware middleware that enables automated optimizations of the application deployment in order to satisfy the application’s functional goals while the execution context changes in terms of available computation, memory and network resources. We use finite state machines to model the reconfiguration of the application based on contextual triggers and facilitate system design through the abstraction of system states. We illustrate the benefits of our approach with an image recognition application with well-defined performance goals that is deployed in a three-tier mobile-edge-cloud architecture.

Food intake monitoring can play an important role in the prevention of malnutrition in the aging population, but traditional tools may not be adequate for use in this target group. These tools typically involve the use of questionnaires or food diaries that require manual data entry. Due to their time-consuming nature, they are often incomplete, contain mistakes, or not used at all. An alternative to self-reporting tools, in the form of a plate system that automatically measures the consumed food during the meal, is presented in this paper. Furthermore, the system can estimate the location where each bite was taken on the plate. The system is compatible with an off-the-shelf plate that is mounted on top of a base station. Weight sensors are integrated in the base, allowing for easy removal and cleaning of the plate. Localization of bites is done by looking at the movement of the center of mass during eating. When used with a compartmentalized plate, the amount of consumed food per compartment can be measured. With prior knowledge of the type of food in each compartment, this can give an indication of calories and nutritional intake. We present a bite detection algorithm using a random forest decision tree classifier. Data from 24 aging adults (ages 52-95) eating a single meal with chopsticks was used to train and evaluate the model. Out of a total of 836 true annotated bites, the algorithm detected 602 with a precision and recall of 0.78 and 0.76, respectively. By summing the weights of detected bites from each compartment, the algorithm was able to estimate the amount of food taken per compartment with an average error of (8 ±8)% of the portion size.

Recommender systems have gained prominence in bringing users tailor-made content from the web aiding their decision making process. However, this personalization comes at a cost of privacy of sharing personal information with the recommendation provider. Moreover, with growing sizes of datasets and models, centralized processing of such data has become challenging. To this end, we propose a federated matrix factorization algorithm to enable personal data to be stored and used on-device for training while sending updates to train a centralized model. We illustrate preliminary results from our algorithm applied to recommendation of articles to users of a mobile application based on their reading history. We compare our performance with centralized matrix factorization applied on the same dataset.

Early gestational weight gain prediction can help expecting women overcome several associated risks. However, training the model requires access to centrally stored privacy sensitive weight and other meta-data. In this demo, we present a privacy preserving federated learning approach where we train a global weight gain prediction model by aggregating client models trained locally on their personal data. We showcase a software data-exploration tool that exhibits local model generation, sharing and updating across users and server for proposed collaborative learning. Our proposed model predicts the final weight category with 61.3% accuracy on day 140, with a 8.8% compromise on the centralized training accuracy.

Today's industry is rapidly evolving towards Industry 4.0 in which the Internet of Things is transforming machines into highly-interactive cyber-physical systems. At its core, these cyber-physical systems are driven by industrial embedded systems. However, the trend of many interacting systems creates a harsher working environment for these embedded systems to operate in, considering among other things electromagnetic interference. This working environment can introduce bit flips in the embedded system's hardware which cause control flow and data flow errors in its control software. This paper analyzes the effects of such control flow errors on an Industry 4.0 setup. We also apply a software implemented control flow error detection method to determine its gain in reliability. Finally, this paper demonstrates a basic but effective automated recovery method for our Industry 4.0 setup.

Due to harsher working environments, soft errors or erroneous bit-flips occur more frequently in microcontrollers during execution. Without mitigation, such errors result in data corruption and control flow errors. Multiple software-implemented mitigation techniques have already been proposed. In this paper, we evaluate seven signature monitoring techniques in seven different test cases. We measure and compare their detection ratios, execution time overhead, and code size overhead. From the gathered results, we derive five requirements to develop an optimal signature monitoring technique. Based on these requirements, we propose a new signature monitoring technique called random additive signature monitoring (RASM). RASM uses signature updates with random values and optimally placed validity checks to detect interblock control flow errors. RASM has a higher detection ratio, lower execution time overhead, and lower code size overhead than the studied techniques.