Abstract Image memorability (IM) estimation typically relies on learning generic semantic features from large-scale datasets; however, memorability is intrinsically individual-dependent and personalized visual viewing behavior, but overlooking this variability can undermine model performance in downstream cognitive and human–machine interaction tasks, leading toward suboptimal performance. To address this, we propose PerMem , a unified framework that integrates knowledge distillation (KD) and imitation learning (IL) to jointly learn generic and personalized salient representations for memorability estimation by leveraging both image content and gaze-derived heatmaps. PerMem employs a teacher network built upon a pre-trained ResNet-50 backbone, followed by an encoder–decoder architecture coupled with spatial and channel attention mechanisms to generate generic saliency-aware memorability maps. A lightweight, attention-guided student encoder–decoder is then optimized through a composite imitation-guided distillation process, where knowledge is distilled from the teacher while simultaneously imitating user-specific gaze fixation heatmaps. Through this joint training process, the student network learns to produce personalized memorability estimates while achieving substantial reductions in computational complexity. We validate PerMem on two public IM estimation datasets (LaMem and SUN) and a in-house WoM dataset comprising of 45 participants (UMBC IRB #670) engaged in visual search and navigation tasks that reflect individualized visual attention patterns. PerMem outperforms nine state-of-the-art IM models by capturing coarse-to-fine saliency and adapting to individual attention, achieving $$\approx $$ ≈ 6% improvement in memorability prediction. Lastly, we evaluate PerMem on heterogeneous embedded edge devices, including Jetson Nano, Jetson Xavier NX, and Raspberry Pi, demonstrating efficient on-device inference with consistent reductions in memory usage (25.6–33.4%), power consumption (19.0–25.0%), and inference latency (34.9–39.8%) relative to the teacher model, highlighting its practical robustness for resource-constrained, human-centric applications.

Developing scalable wearable human activity recognition (wHAR) models is challenging due to domain shifts that substantially degrade performance across downstream tasks. Unsupervised domain adaptation (UDA) seeks to improve generalization by transferring knowledge from labeled source domains to unlabeled target domains. However, conventional UDA methods primarily align marginal feature distributions while neglecting feature-label dependencies, often leading to negative transfer and sub-optimal performance. Motivated by these limitations, we propose a novel optimization framework that tackles two key challenges: (i) generating reliable pseudo-labels for the unlabeled target domain and (ii) minimizing conditional discrepancies across domains. To address (i), we employ temperature-based entropy minimization (TEM), which calibrates prediction confidence by scaling logits with a temperature parameter to produce robust pseudo-labels. For (ii), we introduce a polynomial kernel-based cross-covariance (PkCC) loss, a high-order statistics-driven approach that maps features into a reproducing kernel hilbert space (RKHS) to capture richer feature-label dependencies and reduce conditional distribution gaps between domains. In addition, we demonstrate that CoDAN readily extends to partial UDA (pUDA), where the target label space is a subset of the source, and extensive evaluations on public wHAR datasets with diverse label spaces validate its superior performance over state-of-the-art methods in both UDA and pUDA scenarios.

Image memorability (IM) estimation typically relies on learning generic semantic features from large-scale datasets; however, memorability is intrinsically individual-dependent and personalized visual viewing behavior, but overlooking this variability can undermine model performance in downstream cognitive and human-machine interaction tasks, leading towards suboptimal performance. To address this, we propose MemGaze, a unified framework that integrates knowledge distillation (KD) and imitation learning (IL) to jointly learn generic and personalized salient representations for memorability estimation by leveraging both image content and gaze-derived heatmaps. MemGaze employs a teacher network built upon a pretrained ResNet-50 backbone, followed by an encoder-decoder architecture coupled with spatial and channel attention mechanisms to generate generic saliency-aware memorability maps. A lightweight, attention-guided student encoder-decoder is then optimized through a composite imitation-guided distillation process, where knowledge is distilled from the teacher while simultaneously imitating user-specific gaze fixation heatmaps. Through this joint training process, the student network learns to produce personalized memorability estimates while achieving substantial reductions in computational complexity. We validate MemGaze on two public IM estimation datasets (LaMem and SUN) and a inhouse WoM dataset comprising of 45 participants (UMBC IRB #670) engaged in visual search and navigation tasks that reflect individualized visual attention patterns. MemGaze outperforms nine state-of-the-art IM models by capturing coarse-to-fine saliency and adapting to individual attention, achieving ≈6% improvement in memorability prediction.

Cybercrime is a growing threat to organizations and individuals worldwide, with criminals using sophisticated techniques to breach security systems and steal sensitive data. This paper aims to comprehensively survey the latest advancements in cybercrime prediction, highlighting the relevant research. For this purpose, we reviewed more than 150 research articles and discussed 50 most recent and appropriate ones. We start the review with some standard methods cybercriminals use and then focus on the latest machine and deep learning techniques, which detect anomalous behavior and identify potential threats. We also discuss transfer learning, which allows models trained on one dataset to be adapted for use on another dataset. We then focus on active and reinforcement learning as part of early-stage algorithmic research in cybercrime prediction. Finally, we discuss critical innovations, research gaps, and future research opportunities in Cybercrime prediction. This paper presents a holistic view of cutting-edge developments and publicly available datasets.

Recent advancements in deep learning-based wearable human action recognition (wHAR) have improved the capture and classification of complex motions, but adoption remains limited due to the lack of expert annotations and domain discrepancies from user variations. Limited annotations hinder the model's ability to generalize to out-of-distribution samples. While data augmentation can improve generalizability, unsupervised augmentation techniques must be applied carefully to avoid introducing noise. Unsupervised domain adaptation (UDA) addresses domain discrepancies by aligning conditional distributions with labeled target samples, but vanilla pseudo-labeling can lead to error propagation. To address these challenges, we propose μDAR, a novel joint optimization architecture comprised of three functions: (i) consistency regularizer between augmented samples to improve model classification generalizability, (ii) temporal ensemble for robust pseudo-label generation and (iii) conditional distribution alignment to improve domain generalizability. The temporal ensemble works by aggregating predictions from past epochs to smooth out noisy pseudo-label predictions, which are then used in the conditional distribution alignment module to minimize kernel-based class-wise conditional maximum mean discrepancy (kCMMD) between the source and target feature space to learn a domain invariant embedding. The consistency-regularized augmentations ensure that multiple augmentations of the same sample share the same labels; this results in (a) strong generalization with limited source domain samples and (b) consistent pseudo-label generation in target samples. The novel integration of these three modules in μDAR results in a range of ~ 4-12% average macro-F1 score improvement over six state-of-the-art UDA methods in four benchmark wHAR datasets.

Simulated environments can be a quicker and more flexible alternative to training and testing machine learning models in the real world. Models also need to be able to efficiently communicate with the environment. In military-relevant environments, a trained model can play a valuable role in finding cover for an autonomous robot to prevent getting detected or attacked by adversaries. In this regard, we present a forest simulation and robot control framework that is ready for integration with machine learning or object recognition algorithms. Our framework includes an environment relevant to military situations and is capable of providing information about the environment to a machine learning model. A forest environment was designed with wooded areas, open paths, water, and bridges. A Clearpath Husky robot is simulated in the environment using Army Research Laboratory’s (ARL) Unity and ROS simulation framework. The Husky robot is equipped with a camera and lidar sensor. Data from these sensors can be read through ROS topics and RViz configuration windows. The robot can be moved using ROS velocity command topics. These communication methods can be employed by a machine learning algorithm for use in detecting trees to attain maximum cover. Our designed environment improves upon the default ARL framework environments by offering a more diverse terrain and more opportunities for cover. This makes the environment more relevant to a cover-seeking machine learning model. Code, videos, and integration process available at : https://github.com/avispector7/Forest-Simulation

The COVID-19 pandemic has brought a devastating impact on human health across the globe, and people are still observing face-masking as a preventive measure to contain the spread of COVID-19. Coughing is one of the major transmission mediums of COVID-19, and early cough detection could play a significant r ole i n p reventing t he s pread o f t his life-threatening virus. Many approaches have been proposed for developing systems to detect coughing and other respiratory symptoms in literature, but earable devices are not well-studied and investigated for respiratory symptom detection. In this work, we posited an acoustic research prototype (earable device) - eSense that has acoustic and IMU sensors embedded into user-convenient earbuds to address the following issues: (i) feasibility of the earables in detecting respiratory symptoms, and (ii) scalability of trained machine learning models in the presence of unseen data samples. We performed experimentation with both shallow and deep learning models on the eSense collected data samples. We observed that the deep learning model outperforms the shallow learning models achieving 97% accuracy. Furthermore, we investigated the scalability of the deep learning model on unseen datasets and noticed that the performance of the deep learning model deteriorates when trained on a particular dataset and tested on an unseen dataset. To mitigate such challenges, we postulated an adversarial domain adaptation technique that helps improve the performance of our respiratory symptoms detection framework by a substantial margin.

Machine learning models are bounded by the credibility of ground truth data used for both training and testing. Regardless of the problem domain, this ground truth annotation is objectively manual and tedious as it needs considerable amount of human intervention. With the advent of Active Learning with multiple annotators, the burden can be somewhat mitigated by actively acquiring labels of most informative data instances. However, multiple annotators with varying degrees of expertise poses new set of challenges in terms of quality of the label received and availability of the annotator. Due to limited amount of ground truth information addressing the variabilities of Activity of Daily Living (ADLs), activity recognition models using wearable and mobile devices are still not robust enough for real-world deployment. In this paper, we first propose an active learning combined deep model which updates its network parameters based on the optimization of a joint loss function. We then propose a novel annotator selection model by exploiting the relationships among the users while considering their heterogeneity with respect to their expertise, physical and spatial context. Our proposed model leverages model-free deep reinforcement learning in a partially observable environment setting to capture the action-reward interaction among multiple annotators. Our experiments in real-world settings exhibit that our active deep model converges to optimal accuracy with fewer labeled instances and achieves ~8% improvement in accuracy in fewer iterations.

Following healthy lifestyle is a key for active living. Regular exercise, controlled diet and sound sleep play an invisible role on the well being and independent living of the people. Sleep being the most durative activities of daily living (ADL) has a major synergistic influence on people's mental, physical and cognitive health. Understanding the sleep behavior longitudinally and its underpinning clausal relationships with physiological signals and contexts (such as eye or body movement etc.) horizontally responsible for a sound or disruptive sleep pattern help provide meaningful information for promoting healthy lifestyle and designing appropriate intervention strategy. In this paper we propose to detect the microscopic states of the sleep which fundamentally constitute the components of a good or bad sleeping behavior and help shape the formative assessment of sleep quality. We initially investigate several classification techniques to identify and correlate the relationship of microscopic sleep states with the overall sleep behavior. Subsequently we propose an online algorithm based on change point detection to better process and classify the microscopic sleep states and then test a lightweight version of this algorithm for real time sleep monitoring activity recognition and assessment at scale. For a larger deployment of our proposed model across a community of individuals we propose an active learning based methodology by reducing the effort of ground truth data collection. We evaluate the performance of our proposed algorithms on real data traces, and demonstrate the efficacy of our models for detecting and assessing fine-grained sleep states beyond an individual.

Predicting the occupancy related information in an environment has been investigated to satisfy the myriad requirements of various evolving pervasive, ubiquitous, opportunistic and participatory sensing applications. Infrastructure and ambient sensors based techniques have been leveraged largely to determine the occupancy of an environment incurring a significant deployment and retrofitting costs. In this paper, we advocate an infrastructure-less zero-configuration multimodal smartphone sensor-based techniques to detect fine-grained occupancy information. We propose to exploit opportunistically smartphones' acoustic sensors in presence of human conversation and motion sensors in absence of any conversational data. We develop a novel speaker estimation algorithm based on unsupervised clustering of overlapped and non-overlapped conversational data to determine the number of occupants in a crowded environment. We also design a hybrid approach combining acoustic sensing opportunistically with locomotive model to further improve the occupancy detection accuracy. We evaluate our algorithms in different contexts, conversational, silence and mixed in presence of 10 domestic users. Our experimental results on real-life data traces collected from 10 occupants in natural setting show that using this hybrid approach we can achieve approximately 0.76 error count distance for occupancy detection accuracy on average.

To promote energy-efficient operations in residential and office buildings, non-intrusive load monitoring (NILM) techniques have been proposed to infer the fine-grained power consumption and usage patterns of appliances from power-line measurement data. Fine-grained monitoring of everyday appliances (such as toasters and coffee makers) can not only promote energy-efficient building operations, but also provide unique insights into the context and activities of individuals. Current building-level NILM techniques are unable to identify the consumption characteristics of relatively low-load appliances, whereas smart-plug based solutions incur significant deployment and maintenance costs. In this paper, we investigate an intermediate architecture, where smart circuit breakers provide measurements of aggregate power consumption at room (or section) level granularity. We then investigate techniques to identify the usage and energy consumption of individual appliances from such measurements. We first develop a novel correlation-based approach called CBPA to identify individual appliances based on both their unique transient and steady-state power signatures. While promising, CBPA fails when the set of candidate appliances is too large. To further improve the accuracy of appliance level usage estimation, we then propose a hybrid system called AARPA, which uses mobile sensing to first infer high-level activities of daily living (ADLs), and then uses knowledge of such ADLs to effectively reduce the set of candidate appliances that potentially contribute to the aggregate readings at any point. We evaluate two variants of this algorithm, and show, using real-life data traces gathered from 10 domestic users, that our fusion of mobile and power-line sensing is very promising: it identified all devices that were used in each data trace, and it identified the usage duration and energy consumption of low-load consumer appliances with 87% accuracy.