Compressed learning (CL), integrating compressed sensing (CS) and machine learning (ML), enables direct inference from few CS measurements. However, existing CL methods either heavily rely on pre-trained models built upon extremely large-scale datasets or perform only relatively simple tasks on small datasets, restricting their scalability in real-world multimedia scenarios. To address these limitations, we propose an efficient CL framework named HTMA-CL for practical multimedia acquisition and edge intelligent processing. HTMA-CL employs CNN-based learnable sampling to realize block-based CS for high-resolution images, significantly decreasing the transmission bandwidth and storage overhead. A hierarchical tokenization module together with a deep-narrow Transformer module progressively models local and global dependencies within the measurements, enabling accurate inference directly in the compressive domain. Various task heads are constructed for performing diverse multimedia analysis tasks such as image classification and semantic segmentation. Extensive experiments demonstrate that our HTMA-CL achieves state-of-the-art performance compared to other CL methods, and nearly comparable performance to the image domain methods at a CS ratio of 10%. Our method further verifies the strong robustness against external interference in the disturbance-prone IoT multimedia environments. The source code is publicly available at https://github.com/acrlife/HTMA-CL.git .

Medical object detection plays a vital role in computer-aided diagnosis (CAD), enabling automatic and precise localization of lesions from medical images. In recent years, the YOLO series has been widely adopted as a mainstream real-time detection architecture. However, its commonly used Path Aggregation Feature Pyramid Network (PAFPN) still shows limitations in spatial modeling and fine-grained perception under low-contrast images. Meanwhile, State Space Models (SSMs), with their linear complexity and excellent long-range dependency modeling capability, have demonstrated great potential in both sequential modeling and vision tasks. To this end, this paper proposes a lightweight medical object detection framework dubbed MedMamba-YOLO, which integrates state-space modeling and multi-scale structural awareness. In this framework, we construct a Multi-Scale Fusion FPN (MSF-FPN) and embed a newly designed Spatial-Aware Visual State Space Block (SAVSSB) to achieve efficient cross-scale semantic interaction and contextual modeling. A Cross Path Interaction Block (CPIB) is also devised to strengthen feature representation via cross-path interactions and residual enhancement. Furthermore, a Heterogeneous Multi-scale Dilated Attention Detection Head (HMDADetect) is designed to adaptively model objects at different scales through a multi-branch variable dilation mechanism. Experiments conducted on the Br35H and BCCD datasets demonstrate that MedMamba-YOLO with only 3.8M parameters and 8.8GFLOPs achieves state-of-the-art performance without the need for pretraining. The source codes are publicly available at: https://github.com/HHH-JYG/MedMamba-YOLO.

Establishing an objective and quantifiable analysis of the behaviors of children with autism remains a major challenge in current research. Behavioral analysis is essential for both early autism screening and designing effective intervention strategies. Drawing on theories from psychology and behavioral science, we propose an explainable interactive digital behavioral analysis method grounded in spatial correlation. Using a series of context-based behavioral stimulation paradigms, we collected multi-modal reactive behaviors from 112 children. We then compared the response patterns of children with autism to their clinically observed symptoms to evaluate the developmental status of the core competencies underlying their behaviors. By quantifying children's multi-modal responses through a spatial correlation behavioral paradigm analysis method, we generated digital behavioral representations and applied them to autism screening, achieving an accuracy of 94.4 %. These results validate the effectiveness and reliability of our digital behavioral analysis approach, demonstrating its potential value in supporting objective autism screening and informing subsequent intervention planning.

This paper focuses on the distributed event-triggered adaptive neural resilient time-varying formation control problem for a class of multiple-input multiple-output nonlinear multi-agent systems, where all network communication links between agents are subjected to denial-of-service (DoS) attacks simultaneously. A second-order resilient time-varying formation estimator is designed to obtain the unknown leader information in DoS attack active intervals. Meanwhile, a state-triggering mechanism (STM) is designed to save system communication resources. Nevertheless, the STM can lead to virtual control laws being non-differentiable. To circumvent the problem, we first design an adaptive neural resilient formation control scheme. Then, based on the adaptive neural resilient formation control scheme, we replace continuous states with intermittent ones. By utilizing a dynamic filtering technique, an event-based adaptive neural resilient formation control scheme is designed. The key technology of control scheme design is to establish an improved first-order auxiliary system to deal with the negative impact of actuator saturation. It is proved that formation tracking errors can converge to a residual set around zero, and all signals in the closed-loop system are semi-globally uniformly ultimately bounded. Finally, simulation results are presented to show the effectiveness of the control scheme.

This paper studies the event-triggered time-varying formation control problem for nonlinear multi-agent systems with actuator faults. Based on the neural network approximation technique, a neural observer is constructed to estimate the unmeasured states of systems. Then, a distributed adaptive event-triggered time-varying formation control manner is proposed utilizing the intermittent estimated states information from the agent and its neighbors. To overcome the problem that estimated states triggering leads to virtual control laws is non-differentiable, a distributed continuous control scheme under regular output-feedback is designed firstly, upon which a distributed event-triggered controller is constructed by replacing estimated states with intermittent estimated ones. It is shown that the designed event-triggered output-feedback time-varying formation fault-tolerant controller can compensate for actuator faults, and all signals in closed-loop systems are semi-globally uniformly ultimately bounded. Finally, simulation results of a practical example are given to verify the effectiveness of the proposed control manner.Note to Practitioners—Formation control has broad application prospects in modern military and civilian fields, such as combat aircraft flying formation, satellite formation, autonomous vehicle formation, etc. In formation control systems, when agents occur actuator faults, it may break the original formation and even cause collision between agents. As a result, the security of formation control systems is facing great challenges in practical engineering applications. On the other hand, communication bandwidth is limited in practical engineering systems, and how to make systems quickly form formation under the limited communication bandwidth has become a key topic. Inspired by the above discussions, a distributed state-triggered output-feedback time-varying formation fault-tolerant control scheme is designed in this paper, in which actuator faults are compensated by using adaptive technology. Meanwhile, to sufficiently save the usage of system communication resources, a dual-channel event-triggered mechanism is designed.