AI 中文总结
本文在RISC-V嵌入式平台上评估硬约束循环物理信息模型的KAN残差分支，发现其部署时延迟和能耗高于MLP，量化后可靠性更差，不适合作为默认选择。
AI 中文摘要
硬约束循环物理信息网络（HRPINNs）将已知动力学嵌入循环数值积分器中，仅限制神经网络分支学习第一性原理模型未捕获的残差动力学。柯尔莫哥洛夫-阿诺德网络（KANs）被提出作为多层感知器（MLPs）的参数高效替代方案，用于此类残差分支，但其可学习的B样条激活函数具有明显不同的执行特性。基于先前研究中关于普通B样条KAN作为HRPINN残差分支在发现精度上与MLP相当或表现更差的结论，本文探究该参数效率在部署阶段是否依然存在。使用相同的训练权重，在无向量扩展的RISC-V RV64GC平台（StarFive VisionFive~2，SiFive U74）的闭合循环中，测量执行延迟、每积分步能耗及训练后量化下的可靠性。在两个精度相当的配对中，KAN残差分支的执行速度分别慢13.5倍和8.0倍，每积分步能耗分别增加11.3倍和5.6倍（最小配对中为3.7μJ对比0.33μJ）；在所有四个参数匹配的规模层级中，范围分别为4.7倍至14.5倍和4.7倍至18.7倍。在INT8量化下，KAN的轨迹比匹配的MLP早发散多达43倍；该问题源于权重量化，而非输入侧节点区间分配错误。这些结果表明，KAN所报告的参数效率无法迁移到标量嵌入式核心的部署成本中，除非采用特定的量化协同设计，否则MLP残差分支是嵌入式HRPINN部署更可靠的默认选择。
英文摘要
Hard-constrained recurrent physics-informed networks (HRPINNs) embed known dynamics inside a recurrent numerical integrator and restrict a neural branch to learning only the residual dynamics that the first-principles model does not capture. Kolmogorov--Arnold Networks (KANs) have been proposed as parameter-efficient replacements for multilayer perceptrons (MLPs) in such residual branches, but their learnable B-spline activations follow a markedly different execution profile. Building on prior work that characterized when a vanilla B-spline KAN matches or underperforms an MLP as an HRPINN residual branch in discovery accuracy, this paper asks whether that parameter efficiency survives deployment. Using identical trained weights, we measured execution latency, energy per integration step, and dependability under post-training quantization in the closed recurrent loop on a RISC-V RV64GC platform without vector extensions (StarFive VisionFive~2, SiFive U74). For the two accuracy-comparable pairs, the KAN residual branch executed $13.5\times$ and $8.0\times$ slower and consumed $11.3\times$ and $5.6\times$ more energy per integration step (3.7\,$μ$J against 0.33\,$μ$J for the smallest pair); across all four parameter-matched size tiers the ranges are $4.7\times$--$14.5\times$ and $4.7\times$--$18.7\times$. Under INT8 quantization, KAN trajectories diverged up to $43\times$ earlier than matched MLPs; the damage traces to weight quantization, not to input-side knot-interval misassignment. These results indicate that the parameter efficiency reported for KANs does not transfer to deployment cost on scalar embedded cores, and that an MLP residual branch is the more dependable default for embedded HRPINN deployment unless specific quantization co-design is used.
Comments6 pages, submitted to SBESC 2026 as an extension to the ICLR Workshop Paper https://openreview.net/forum?id=iHpbQn99RB
