Task-Directed Residual AddUNet:Perfect-Reconstruction Routing for Full-Rate Representations
本文提出了一种任务导向的残差全速率完美重构架构,通过路由无关紧要或冗余信息来优化面向任务的表现,同时保证精确重构。
本文提出了一种任务导向的残差全速率完美重构架构,通过路由无关紧要或冗余信息来优化面向任务的表现,同时保证精确重构。
本文提出AKRASIA,一种针对基于推理的代码语言模型的隐秘后门攻击方法,通过构建代码级触发器和利用上下文学习来实现恶意目标并规避检测。
本文针对GAN生成的假脸检测难题,提出结合EfficientNet与Swin Transformer的轻量级架构方法,提高了检测准确率和效率。
研究通过在浮点乘法器中采用部分积矩阵列截断和正负压缩器集成的方法,优化了计算效率与精度之间的权衡,实现了硬件加速器的显著改进。
This study addresses performance degradation in critical ROS 2 nodes caused by backpressure from slow subscribers on reliable DDS topics. We propose a proxy-based decoupling architecture featuring a novel dual-write mechanism and a health probe classifier with hysteresis. By splitting topics and isolating traffic via parallel reliable and best-effort channels, the system adaptively throttles non-critical data streams to prevent backpressure propagation. Experimental results demonstrate that under severe interference, publisher throughput remains stable at 30 Hz, while the p95 tail latency for critical subscribers decreases from 15 seconds to under 2 milliseconds. These findings confirm that the proposed approach effectively preserves both system real-time responsiveness and overall throughput in the presence of heterogeneous subscriber behaviors.
本文提出了一种任务导向的残差全速率完美重构架构,通过路由无关紧要或冗余信息来优化面向任务的表现,同时保证精确重构。
本文提出AKRASIA,一种针对基于推理的代码语言模型的隐秘后门攻击方法,通过构建代码级触发器和利用上下文学习来实现恶意目标并规避检测。
本文针对GAN生成的假脸检测难题,提出结合EfficientNet与Swin Transformer的轻量级架构方法,提高了检测准确率和效率。
研究通过在浮点乘法器中采用部分积矩阵列截断和正负压缩器集成的方法,优化了计算效率与精度之间的权衡,实现了硬件加速器的显著改进。
This study addresses performance degradation in critical ROS 2 nodes caused by backpressure from slow subscribers on reliable DDS topics. We propose a proxy-based decoupling architecture featuring a novel dual-write mechanism and a health probe classifier with hysteresis. By splitting topics and isolating traffic via parallel reliable and best-effort channels, the system adaptively throttles non-critical data streams to prevent backpressure propagation. Experimental results demonstrate that under severe interference, publisher throughput remains stable at 30 Hz, while the p95 tail latency for critical subscribers decreases from 15 seconds to under 2 milliseconds. These findings confirm that the proposed approach effectively preserves both system real-time responsiveness and overall throughput in the presence of heterogeneous subscriber behaviors.