X-RACE: XAI-assisted Recurrent neural network Attribution for Channel Estimation
本文提出X-RACE框架，通过低复杂度的一次性双优化策略解决高移动车辆环境下的信道估计问题，同时评估和修剪无关输入子载波和内部隐藏单元。
0 citationsRead paper
Selected work
Representative Papers
3D Weighted Geometric Graph Neural Networks for Sheep Facial Pain Assessment
Traditional 2D approaches struggle to model the three-dimensional anatomical structure of sheep faces and the spatial relationships among facial keypoints, limiting the accuracy of Sheep Pain Facial Expression Scale (SPFES)-based pain assessment. This work proposes 3D-SPFES, a novel system that integrates monocular RGB-based depth estimation with a weighted geometric graph neural network (WG-GNN). Leveraging VideoDepthAnything to recover depth, the method embeds facial keypoints into 3D Euclidean space and constructs edge weights that combine Euclidean distances with surface coplanarity. An anatomy-aware scaled dot-product attention mechanism is introduced to facilitate geometric message passing, ultimately yielding a continuous, normalized pain score ranging from 0% to 100%. Requiring no specialized depth-sensing hardware, this approach substantially enhances the modeling of SPFES-relevant features and their 3D spatial dependencies, enabling highly accurate automated pain assessment.
0 citationsRead paper
Price Information Is Not Enough: Ordering and Decision Rules in Storage Bidding
This study addresses the limitation of relying solely on price forecasting for optimizing energy storage bidding, as actual revenue depends critically on the interaction between decision rules and price signals. The work disentangles the theoretical value of price signals from the gains achievable through strategic implementation and proposes an ordinal bidding strategy that leverages only the ranking of prices—rather than their absolute accuracy—under daily throughput constraints. Through a conditional sub-Gaussian price model, mutual information analysis, and Gaussian perturbation experiments, the authors reveal a non-monotonic relationship between information content and realized revenue. Empirical results demonstrate that the ranking-based strategy captures 90% of the attainable revenue; even with perfectly accurate price vectors, swapping the highest and lowest prices reduces revenue by 53%; and current climate-aware benchmarks already achieve 78% of the revenue attainable under ideal forecasts.
0 citationsRead paper
Anonymous sharing is pairwise phase-blind
This study addresses the “checkpoint storm” problem in shared-storage systems, where concurrent checkpoint writes from multiple training tasks cause severe I/O contention. The authors model training tasks as pulse-coupled oscillators interacting through an anonymous shared resource. Drawing on nonlinear dynamical systems theory, phase map analysis, and extensions of the Kuramoto and Mirollo–Strogatz models—supported by numerical simulations—they demonstrate that under anonymous resource sharing, inter-task phase coupling does not emerge. Consequently, the synchronous state is not an attractor but an unstable fixed point with multiple expanding directions. The analysis proves that anonymous sharing alone cannot induce phase locking or clustering. Although tail-end concurrent write loads remain higher than in fully asynchronous scenarios, genuine coupling arises only when resources are constrained and tasks exhibit heterogeneity.
0 citationsRead paper
Do Co-Located AI Training Jobs Synchronize? Load-Dependent Throttling as a Coupling Mechanism for Phase-Locking Behind a Shared Power Cap
This work investigates emergent synchronization among multiple independent AI training tasks operating under a shared power budget, where load-dependent throttling mechanisms—such as power capping, voltage droop, and shared cooling—can induce collective dynamics that cause aggregate power fluctuations to scale linearly rather than decay with the square root of the number of tasks as conventionally expected. By modeling colocated training workloads as a generalized Kuramoto system, the study reveals for the first time that power management mechanisms alone can drive phase locking even in the absence of explicit clock synchronization. Theoretical analysis demonstrates that when phase lags satisfy specific conditions, the effective coupling transitions from repulsive to attractive, leading to a first-order synchronization transition with hysteresis. Building on these insights, the authors propose a phase-scattering scheduling strategy that raises the synchronization threshold, validated through dual-task power-capping experiments aligning closely with theoretical predictions.
0 citationsRead paper
Recent publications
Latest Papers
X-RACE: XAI-assisted Recurrent neural network Attribution for Channel Estimation
本文提出X-RACE框架，通过低复杂度的一次性双优化策略解决高移动车辆环境下的信道估计问题，同时评估和修剪无关输入子载波和内部隐藏单元。
0 citationsRead paper
3D Weighted Geometric Graph Neural Networks for Sheep Facial Pain Assessment
Traditional 2D approaches struggle to model the three-dimensional anatomical structure of sheep faces and the spatial relationships among facial keypoints, limiting the accuracy of Sheep Pain Facial Expression Scale (SPFES)-based pain assessment. This work proposes 3D-SPFES, a novel system that integrates monocular RGB-based depth estimation with a weighted geometric graph neural network (WG-GNN). Leveraging VideoDepthAnything to recover depth, the method embeds facial keypoints into 3D Euclidean space and constructs edge weights that combine Euclidean distances with surface coplanarity. An anatomy-aware scaled dot-product attention mechanism is introduced to facilitate geometric message passing, ultimately yielding a continuous, normalized pain score ranging from 0% to 100%. Requiring no specialized depth-sensing hardware, this approach substantially enhances the modeling of SPFES-relevant features and their 3D spatial dependencies, enabling highly accurate automated pain assessment.
0 citationsRead paper
Price Information Is Not Enough: Ordering and Decision Rules in Storage Bidding
This study addresses the limitation of relying solely on price forecasting for optimizing energy storage bidding, as actual revenue depends critically on the interaction between decision rules and price signals. The work disentangles the theoretical value of price signals from the gains achievable through strategic implementation and proposes an ordinal bidding strategy that leverages only the ranking of prices—rather than their absolute accuracy—under daily throughput constraints. Through a conditional sub-Gaussian price model, mutual information analysis, and Gaussian perturbation experiments, the authors reveal a non-monotonic relationship between information content and realized revenue. Empirical results demonstrate that the ranking-based strategy captures 90% of the attainable revenue; even with perfectly accurate price vectors, swapping the highest and lowest prices reduces revenue by 53%; and current climate-aware benchmarks already achieve 78% of the revenue attainable under ideal forecasts.
0 citationsRead paper
Anonymous sharing is pairwise phase-blind
This study addresses the “checkpoint storm” problem in shared-storage systems, where concurrent checkpoint writes from multiple training tasks cause severe I/O contention. The authors model training tasks as pulse-coupled oscillators interacting through an anonymous shared resource. Drawing on nonlinear dynamical systems theory, phase map analysis, and extensions of the Kuramoto and Mirollo–Strogatz models—supported by numerical simulations—they demonstrate that under anonymous resource sharing, inter-task phase coupling does not emerge. Consequently, the synchronous state is not an attractor but an unstable fixed point with multiple expanding directions. The analysis proves that anonymous sharing alone cannot induce phase locking or clustering. Although tail-end concurrent write loads remain higher than in fully asynchronous scenarios, genuine coupling arises only when resources are constrained and tasks exhibit heterogeneity.
0 citationsRead paper
Do Co-Located AI Training Jobs Synchronize? Load-Dependent Throttling as a Coupling Mechanism for Phase-Locking Behind a Shared Power Cap
This work investigates emergent synchronization among multiple independent AI training tasks operating under a shared power budget, where load-dependent throttling mechanisms—such as power capping, voltage droop, and shared cooling—can induce collective dynamics that cause aggregate power fluctuations to scale linearly rather than decay with the square root of the number of tasks as conventionally expected. By modeling colocated training workloads as a generalized Kuramoto system, the study reveals for the first time that power management mechanisms alone can drive phase locking even in the absence of explicit clock synchronization. Theoretical analysis demonstrates that when phase lags satisfy specific conditions, the effective coupling transitions from repulsive to attractive, leading to a first-order synchronization transition with hysteresis. Building on these insights, the authors propose a phase-scattering scheduling strategy that raises the synchronization threshold, validated through dual-task power-capping experiments aligning closely with theoretical predictions.
0 citationsRead paper