Human-AI collaborative decision-making can surpass the performance of either humans or AI alone by leveraging their complementarity. However, most existing studies assume collaboration between either an individual or a group of humans and a single AI, leading to inflexible team compositions that hinder the formation of optimal human-AI teams under task constraints. Therefore, a promising direction is to adaptively construct teams by selecting specific numbers of humans and AI from candidate pools under task constraints. There are two challenges: (1) How to model typical yet heterogeneous task constraints? (2) How to efficiently identify the human-AI team within a large solution space? To solve these challenges, we propose TaskHAI, a complementarity-driven adaptive selection strategy for task-constrained human-AI teaming. Specifically, we propose a unified multi-step selection framework that models two types of constraints: budget and minimum confidence, and an adaptive selection strategy that incrementally adds human and AI participants by maximizing Bayesian complementarity, thereby efficiently constructing human–AI teams. Experiments on three image classification datasets show that our method surpasses other human-AI collaboration methods, improving average accuracy by 1.41
Commodity Wi-Fi offers a promising platform for pervasive, non-contact home vital signal monitoring. Among the vital signals accessible through Wi-Fi Channel State Information (CSI), respiration is the most fundamental, underpinning applications from respiratory disorder screening to whole-night sleep staging. However, reliable Wi-Fi-based monitoring in unconstrained home environments faces three progressive challenges: (1) spatial heterogeneity of respiratory modulation across changing postures and locations; (2) temporal non-stationarity of signal quality within overnight recordings, where motion-corrupted episodes collapse monolithic processing models; and (3) scarcity of labeled data for training downstream tasks such as sleep staging. These challenges call for a system that adapts its sensing and inference to changing spatial and signal conditions. We propose Pervital, a scenario-adaptive system with a layered design: a spatial-adaptive waveform construction pipeline that tracks the optimal sensing axis and applies phase-aligned subcarrier fusion; a signal quality-aware network coupling a quality-adaptive encoder with a Mixture-of-Experts decoder for graceful degradation during corrupted episodes; and a sleep rhythm-guided contrastive regularization strategy that exploits intra-night and cross-night sleep structure to amplify training signal from limited recordings. Evaluated on comprehensive LoS/NLoS scenarios and 74 full-night recordings in authentic home environments, Pervital achieves sub-1 bpm respiration error and 76.5
In cross-device collaboration, federated learning (FL) with Low-Rank Adaptation (LoRA) facilitates the efficient adaptation of foundation models. One-shot FL (OFL) has emerged as a compelling alternative to multi-round FL, aiming to reduce communication overhead. However, due to task heterogeneity, OFL often exacerbates knowledge forgetting during model aggregation. Existing studies leverage Mixture-of-Experts (MoE) to decrease forgetting. Nevertheless, OFL-MoE treats all participating nodes as independent experts, which creates a new challenge: the massive number of experts significantly increases the difficulty of MoE network training, leading to a decline in performance. Therefore, this paper proposes a novel OFL-MoE framework based on dynamic clustering (DCoFL) to alleviate this challenge. Specifically, DCoFL first leverages the MoE network and a clusterer to synthesize data in an adversarial manner. Then, it re-clusters the experts using expert features and routing scores. Finally, DCoFL employs Cluster-level loss and intra-cluster variance constraints to enhance the quality of the MoE network. This method not only reduces the difficulty of MoE network training but also improves the efficiency of knowledge fusion. The DCoFL scheme can be used in any OFL paradigm, including traditional heterogeneous networks and LoRA. Through comprehensive experiments on eight datasets, DCoFL improves the classification accuracy by approximately 1
Machine learning models are increasingly deployed in IoT, pervasive computing, and intelligent systems operating under dynamic conditions. While numerous uncertainty quantification (UQ) techniques have been proposed, their comparative behavior under realistic deployment shifts remains insufficiently understood, particularly in IoT environments. This paper presents a systematic study of representative UQ techniques in the context of Human Activity Recognition (HAR), a domain characterized by frequent shifts such as unseen activities, unseen users, and cross-dataset deployments. Rather than introducing a new uncertainty estimation method, our objective is to categorize, adapt, calibrate, and compare techniques that observe uncertainty at different levels of the inference pipeline. We introduce a practical taxonomy encompassing input-space, latent-space, output-space, prediction-space, and internal-dynamics-space uncertainty signals, and evaluate a diverse set of methods using a Transformer-based HAR architecture and a multi-shift experimental protocol. The results show that no uncertainty signal is universally superior: output- and prediction-space methods are most informative under familiar conditions and moderate user shifts, whereas input-, reconstruction-, and internal-dynamics-based signals are more sensitive to unseen activities and severe cross-dataset shifts. We further study how continuous uncertainty scores can be transformed into actionable reliability indicators through threshold calibration. Finally, we show that cascade-based monitoring should be configured according to the expected shift type, coverage requirements, and application risk, rather than treated as a universal aggregation strategy.
Mixed Reality (MR), which includes both Augmented Reality (AR) and Virtual Reality (VR), has been increasingly used to support communication, collaboration, learning, and professional practices. This study applies a computational topic modeling approach to 4,775 Scopus-indexed publications from 1995 to 2024, retrieved on 9 October 2024 using predefined MR and communication-related search terms and exclusion criteria. Using Latent Dirichlet Allocation (LDA), we identified six major topics: design, social, application, training, user, and learning. Within the retrieved corpus, design and user had the largest proportional shares in 1995–1997. Compared with that period, social, learning, and training had higher shares in 2022–2024, while design, user, and application had lower shares. Coverage of the final period is incomplete because of the retrieval cutoff. Beyond topic classification, this study proposes a conceptual organizing model to interpret thematic relationships among the six topics. The findings provide a long-term thematic map of the retrieved literature and offer a basis for future empirical and theory-driven studies on how MR technologies mediate interaction, coordination, feedback, and shared understanding. These findings are specific to the corpus defined by the search strategy and may not generalize to all MR communication research.
