Towards a Connected Heterogeneous All-Medium Integrated Network (CHAIN) for Converged Connectivity Across Land, Sea, Air, and Space
论文提出CHAIN框架,通过全光网络骨干和AI驱动的跨域控制层整合海陆空天各领域通信,解决下一代连接中的跨介质问题。
论文提出CHAIN框架,通过全光网络骨干和AI驱动的跨域控制层整合海陆空天各领域通信,解决下一代连接中的跨介质问题。
为解决多代理LLM工作流中的成本和质量权衡问题,提出ProgRouter框架,通过在线进度引导选择LLM代理,平衡任务进展、时间和成本。
Cross-modal heterogeneity poses a significant challenge to the effective integration of genomic and neuroimaging data, thereby limiting precise diagnosis of neurological disorders. To address this issue, this work proposes GeneFuse, a novel multimodal learning framework that, for the first time, incorporates a pretrained genomic language model (GLM) into imaging-genomics fusion tasks. GeneFuse preserves sequence context through genotype-conditioned feature modulation (GCFM) and dynamically adjusts the contribution of genomic features via an uncertainty-aware residual fusion mechanism. Evaluated under APOE-stratified scenarios, the proposed method achieves AUROC scores of 0.77 and 0.83 on NC vs. MCI and NC vs. AD classification tasks, respectively, substantially outperforming existing fusion approaches.
This work addresses the high energy consumption and data-movement bottlenecks confronting AI deployment in edge and distributed settings by proposing a digitally orchestrated hybrid computing architecture that synergistically integrates analog and neuromorphic computing. Physical computing units are selectively introduced only where they offer substantial energy-efficiency gains, while a unified digital control layer manages uncertainty and ensures fault tolerance. Departing from conventional peak TOPS/W metrics, the study establishes a deployment-oriented, system-level efficiency evaluation framework. By combining photonic computing, in-memory computing, and neuromorphic hardware with programmable digital control, mature software stacks, and comprehensive system-level energy accounting, the work delineates the architecture’s suitability for matrix operations and event-driven tasks, systematically uncovering its energy-efficiency potential, software requirements, and engineering challenges to provide both a theoretical foundation and practical pathway for efficient AI system design.
This study addresses the inefficiency of manual REBA assessments in industrial settings and the privacy concerns associated with vision-based approaches by proposing, for the first time, an end-to-end multitask learning framework leveraging millimeter-wave radar to enable privacy-preserving automatic REBA scoring through 3D human skeletal reconstruction. The method integrates biomechanical constraints and temporal smoothness losses, and employs an oversampling strategy to mitigate data imbalance in high-risk postures. Evaluated on the MMFi dataset, the model achieves a REBA risk-level classification accuracy of 77.78%, with a mean absolute error of 0.93 for high-risk samples and an inference latency of only 5.70 milliseconds.
论文提出CHAIN框架,通过全光网络骨干和AI驱动的跨域控制层整合海陆空天各领域通信,解决下一代连接中的跨介质问题。
为解决多代理LLM工作流中的成本和质量权衡问题,提出ProgRouter框架,通过在线进度引导选择LLM代理,平衡任务进展、时间和成本。
Cross-modal heterogeneity poses a significant challenge to the effective integration of genomic and neuroimaging data, thereby limiting precise diagnosis of neurological disorders. To address this issue, this work proposes GeneFuse, a novel multimodal learning framework that, for the first time, incorporates a pretrained genomic language model (GLM) into imaging-genomics fusion tasks. GeneFuse preserves sequence context through genotype-conditioned feature modulation (GCFM) and dynamically adjusts the contribution of genomic features via an uncertainty-aware residual fusion mechanism. Evaluated under APOE-stratified scenarios, the proposed method achieves AUROC scores of 0.77 and 0.83 on NC vs. MCI and NC vs. AD classification tasks, respectively, substantially outperforming existing fusion approaches.
This work addresses the high energy consumption and data-movement bottlenecks confronting AI deployment in edge and distributed settings by proposing a digitally orchestrated hybrid computing architecture that synergistically integrates analog and neuromorphic computing. Physical computing units are selectively introduced only where they offer substantial energy-efficiency gains, while a unified digital control layer manages uncertainty and ensures fault tolerance. Departing from conventional peak TOPS/W metrics, the study establishes a deployment-oriented, system-level efficiency evaluation framework. By combining photonic computing, in-memory computing, and neuromorphic hardware with programmable digital control, mature software stacks, and comprehensive system-level energy accounting, the work delineates the architecture’s suitability for matrix operations and event-driven tasks, systematically uncovering its energy-efficiency potential, software requirements, and engineering challenges to provide both a theoretical foundation and practical pathway for efficient AI system design.
This study addresses the inefficiency of manual REBA assessments in industrial settings and the privacy concerns associated with vision-based approaches by proposing, for the first time, an end-to-end multitask learning framework leveraging millimeter-wave radar to enable privacy-preserving automatic REBA scoring through 3D human skeletal reconstruction. The method integrates biomechanical constraints and temporal smoothness losses, and employs an oversampling strategy to mitigate data imbalance in high-risk postures. Evaluated on the MMFi dataset, the model achieves a REBA risk-level classification accuracy of 77.78%, with a mean absolute error of 0.93 for high-risk samples and an inference latency of only 5.70 milliseconds.