Yadang Chen

Wearable consumer electronics integrate heterogeneous physiological, behavioural, and contextual signals, making reliable feature-interaction modelling essential when input availability varies across sensing and access conditions. This challenge is especially relevant to women-specific wearable applications, where menstrual-health analysis must extract informative patterns from longitudinal physiological and rhythm measurements. Here, we develop HQxDeepFM, a role-aligned hybrid quantum–classical architecture that assigns shallow variational quantum circuits to the low-order, explicit high-order, and implicit nonlinear components of xDeepFM. Group-wise feature gating characterises predictive behaviour under predefined sensor- and policy-related availability conditions, while multilayer-perceptron-assisted zero-noise extrapolation evaluates quantum expectation recovery under controlled density-matrix channels. The framework is evaluated across three public wearable classification tasks covering menstrual-cycle stages, affective states, and contextual activities, using participant-, subject-, or user-disjoint validation with fold-level confidence intervals and multiplicity-adjusted statistical comparisons. Across these tasks, HQxDeepFM and HQxDeepFMEM exhibit a favourable mean-performance profile relative to the evaluated baselines, while the statistical analysis reveals task-dependent variation in effect magnitude, fold-level consistency, and the strength of the comparative evidence. Ablation and capacity analyses further indicate that coordinated branch specialisation contributes more consistently than increasing circuit width or depth alone. Feature-withholding experiments trace how predictive retention changes as predefined input groups become unavailable, while the channel analysis identifies distinct recovery profiles across channel types and circuit widths. Together, these findings position role-aligned quantum expectation features as a compact design strategy for heterogeneous wearable feature-interaction learning and provide a focused methodological basis for broader women-specific evaluation, repeated-run inference, explicit privacy mechanisms, and calibrated-device studies.

Point cloud registration aligns observations of objects and scenes, yet repeated structures leave a correspondence-selection problem unresolved: semantic grouping can reduce matching ambiguity while removing spatial support needed to constrain rigid motion, making label accuracy an incomplete guide to registration quality. A semantic–spatial framework, SSAR, addresses this problem by linking multiscale contextual descriptors, semantic candidate groups, and soft-correspondence-based rigid estimation, with analytical conditions that relate preferential retention of true matches and the geometry of retained points to pose sensitivity. The empirical analysis draws on recorded baseline experiments using Bunny for neighborhood description, ShapeNet Part for object decomposition and alignment, and Stanford Large-Scale 3D Indoor Spaces (S3DIS) for indoor-scene structure, together with positional-encoding comparisons, feature-learning histories, and runtime measurements under controlled experimental settings. The multiscale pathway yields higher descriptor recall and decomposition scores within the reported comparisons, while the analysis shows that increasing the inlier fraction and maintaining noncollinear support are complementary requirements, linking the composition of the candidate set to the conditioning of rotation estimation. These findings provide a basis for assessing semantic decomposition through the correspondence information it preserves and for designing local registration procedures that jointly consider matching ambiguity, geometric support, and the consistency of retained structures with a common rigid transformation.

Linux kernel specialization can improve system efficiency, but configuration changes may also alter functionality, compatibility, and security. Performance gains are difficult to interpret unless the configuration that was realized and executed is established. Existing configuration-tailoring, performance-search, and LLM-assisted methods address parts of this problem separately. They do not connect workload-derived evidence, model proposals, deterministic Kconfig realization, booted-state identity, empirical performance, and deployment-policy compliance in one auditable process. Intelligent Kernel Configuration (IKC) addresses this gap through an evidence-bounded workflow. Workload traces guide candidate selection, and local Kconfig knowledge constrains LLM-assisted editing. Deterministic configuration tools establish realizable states, while boot-time configurations define experimental identity. Repeated measurements provide performance evidence, and an explicit policy gate determines deployment eligibility. Across representative web-serving, database, storage, media-processing, and scheduler workloads, IKC produced workload-dependent improvements over conventional configurations. Several workload classes retained supported advantages over machine-tailored alternatives. Evidence for the contribution of trace conditioning was more concentrated, and explicit hardening constraints preserved much of the measured performance. Complementary network-path and performance-oriented comparisons were consistent with this interpretation. These findings position model assistance within a traceable realization and validation framework, linking generated recommendations to explicit system evidence and operational outcomes. The resulting workflow establishes a reproducible foundation for workload-aware kernel specialization under well-defined operational constraints.

Quantum neural networks (QNNs) support quantum-state classification and finite-size phase recognition, but expressive parameterized circuits can attenuate gradients and compress class-dependent information at measured readouts. Most mitigation strategies modify circuit design or training schedules. Whether an objective-level geometric constraint can preserve readout separability under practical measurement conditions remains unclear. We introduce a trace-margin quantum neural network (TMQNN) that augments cross-entropy with a local readout trace-distance margin between class-mean reduced states on measurable subsystems. The method retains a shared parameterized circuit. We evaluate TMQNN on public quantum-data benchmarks, including TensorFlow Quantum transverse-field Ising and XXZ finite-size regime-recognition tasks and a QDataSet-derived ternary noise-level classification task constructed from quantum-evolution records. Across these settings, TMQNN yields more structured readout geometry, more stable learning trajectories, less compressed empirical quantum neural tangent kernel profiles, coherent ablation behavior, and retained effectiveness under representative noise channels and finite-shot measurements. The results indicate that local quantum distinguishability can provide a geometric training signal for measurable QNN representations under near-term measurement constraints.

Women-centric wearable and phone/watch consumer-electronics services require privacy-bounded decision layers that integrate physiological rhythms, device states, contextual cues, service history, and user-authorized feature availability, while maintaining compactness and reliability for adaptive support. HQxDeepFM is developed as a privacy- and noise-aware hybrid quantum–classical feature-interaction engine that preserves the functional decomposition of xDeepFM by replacing its Linear, CIN, and DNN branches with QLinear, QCIN, and QDNN modules, where shallow variational quantum circuits generate measured Pauli-Z expectation features for low-order effects, explicit cross-field interactions, and implicit nonlinear representations. Experimental evaluation is conducted on three public wearable and phone/watch datasets: the Menstrual Sleep-Wake (MSW) dataset for smartwatch-assisted menstrual-cycle service adaptation, WESAD for transferable affective well-being support, and ExtraSensory for context-recognition primitives under participant-, subject-, or user-disjoint validation, policy-defined feature-availability masking, missing-sensor robustness tests, 2–8-qubit capacity analysis, and controlled phase-damping, amplitude-damping, and depolarizing noise-channel evaluation. Across the evaluated tasks, HQxDeepFM improves the accuracy–compactness trade-off relative to sparse-interaction baselines, while HQxDeepFMEM, its error-mitigated variant, provides consistent additional gains and high post-mitigation AUC retention under NISQ-relevant perturbations. These results indicate that compact, privacy-bounded, and NISQ-aware feature-interaction learning can support responsible decision-layer design for women-centric consumer electronics, with practical relevance to menstrual-health support, affective well-being services, context-aware safety assistance, and future user-centered wearable intelligence.

With the rapid development of deep learning for HSI–LiDAR land-cover classification, Mamba has shown strong potential in feature representation and generalization. However, insufficient semantic interaction among multimodal features and limited multi-scale feature modeling still restrict classification performance. To address these challenges, we propose a novel Multi-scale Convolution and Cross-attention Guided Mamba (MCCGM) network for joint HSI-LiDAR classification. This framework fully exploits the complementary strengths of distinct architectural components. Specifically, we design a Morphological-feature integrated Multi-scale CNN (MMC) to extract spatial–spectral features from HSI, elevation information from LiDAR, and their morphological features, thereby enabling precise characterization of local edges and texture details of ground objects. In parallel, the Mamba Fusion Encoder (MFE) efficiently models long-range spectral dependencies to promote deeper interaction between HSI spectral features and LiDAR-derived structural information. To further enhance semantic alignment across modalities, a CNN-enhanced cross-attention mechanism is embedded into the Mamba block, where multi-scale convolutions generate enriched queries, keys, and values to highlight modality-specific correlations. The resulting attention-refined tokens are subsequently processed by Mamba’s state-space modeling, which strengthens global dependency learning while preserving inter-modality alignment. Extensive experiments conducted on four publicly available benchmark datasets demonstrate that MCCGM consistently outperforms state-of-the-art algorithms, achieving superior classification accuracy and detection performance. The source code is publicly available at https://github.com/cupid6868/MCCGM.

Although few-shot segmentation (FSS) methods have achieved remarkable results, there remain challenges associated with the limited number of support samples. i) The objects in support and query images may have substantially different appearances even though they belong to the same category, which is known as the prototype bias problem. ii) Most methods neglect the background information, especially the query background during the inference stage. To address these problems, we propose DPRNet, a novel network with dual branch of foreground and background prototype refinement modules. Specifically, we first present a Variational Feature Semantic Enhancement (VFSE) module, in which we refine the object prototype with a variational autoencoder and word-text labels. In this way, the biased class-wise prototype caused by the limited support samples can be aligned, achieving better performance. Second, we design a Background Prototype Refinement (BPR) module that effectively explores the potential information in the background for both the support and query images. More importantly, it is designed to generate online predictions of the query background during the training stage to fully mimic the inference stage. These advancements enhance the robustness and generalizability of our method, and the results of experiments demonstrate its effectiveness. In the 1-way 5-shot setting on PASCAL-$5^{i}$, our method achieves a mean-IoU improvement of 1.59% over the competing method.