ID 原文 译文
56308 为解决这一问题, 本项工作提出面向光流的 CNN 硬件加速设计方案 (称为 Swan-AOE), 即通过支持卷积、反卷积、双线性插值和关联操作解决这类神经网络的硬件加速计算问题. Swan-AOE provides aconfigurable architecture together with an adaptive schedule that offers multiple types of parallelism to achieve theoptimal throughput.
56309 Swan-AOE 包括可配置的硬件计算架构和自适应的调度策略, 通过提供灵活的并行调度实现最优化吞吐量计算.此外, Swan-AOE 还进行设计空间探索, 探索可用片上缓存资源在提高能耗 – 面积效率的潜在能力. In addition, Swan-AOE introduces a design space exploration of the potential of availableon-chip memory resources for energy-area efficiency.
56310 实验结果表明, 与基准加速器相比, 所提出的设计能有效提升性能、能效和面积效率 Experimental results show that the proposed design canefficiently improve the performance, energy efficiency, and area efficiency compared with a comparable combinedaccelerator baseline.
56311 针对五次间接PH曲线的判别问题,本文结合高斯消元法与几何方法给出Bézier控制多边形满足的充分必要条件. This paper studies the problem of identification of quintic indirect-PH curves.
56312 间接PH曲线通过一个二次有理参数变换后,其等距线是有理形式的. By employing Gaussianelimination and geometric approaches, we give necessary and sufficient conditions for a planar parametric curveto be an indirect-PH curve. Indirect-PH curves own rational offsets after being reparameterized by a fractionalquadratic transformation.
56313 间接PH曲线的代数充分必要条件本质是其一阶导数的因式分解满足特定条件,是一种积的形式. Algebraic conditions for curves to have rational offsets are constraints on their firstderivatives, which are the product of polynomials.
56314 考虑到Bézier曲线的表示是Bernstein多项式形式,是一种和的形式. However, B´ezier curves are represented as a sum of Bernsteinpolynomials. By considering the equivalence of the product and the sum, we derive non-linear systems anddiscuss the solutions.
56315 通过这两种形式的相容性引出待求解的非线性方程组并讨论求解问题,最后将所得结果应用在控制多边形上,得到五次间接PH曲线的几何特征. Finally, the results are applied to study the B´ezier control polygon, thus we get geometriccharacteristics of quintic indirect-PH curves.
56316 门控循环单元(gated recurrent unit, GRU)是一种有代表性的深度神经网络,它在众多序列学习任务中达到了国际领先的水平. A gated recurrent unit (GRU) is a representative deep neural network that has achieved promisingresults in many sequence learning tasks.
56317 然而,在门控循环单元的每个时间步之间,输入信息与隐含状态信息缺乏交互,这对更好地挖掘上下文语义信息带来了挑战. However, there is a lack of interaction between the input and the hidden?state among each time step of GRU, resulting in the challenges to mine contextual semantic information effectively.