RISC-V Evaluation of KANs in Hard-Constrained Recurrent Physics-Informed Models
A recent study published on arXiv (2608.00737) examines the use of Kolmogorov-Arnold Networks (KANs) as residual components within hard-constrained recurrent physics-informed networks (HRPINNs) on a RISC-V platform. This research builds upon earlier comparisons between KANs and multilayer perceptrons (MLPs) regarding discovery accuracy, now focusing on whether the parameter efficiency of KANs leads to improved execution performance. The authors evaluated execution latency, energy consumption per integration step, and reliability during post-training quantization using the same trained weights in a closed recurrent loop. Experiments were carried out on a StarFive VisionF, a RISC-V RV64GC platform lacking vector extensions. This cross-type announcement highlights findings significant for edge computing and embedded AI applications facing resource limitations.
Key facts
- Paper arXiv:2608.00737 evaluates KANs in HRPINNs on RISC-V.
- Study measures execution latency, energy per integration step, and dependability under post-training quantization.
- Experiments conducted on StarFive VisionF, a RISC-V RV64GC platform without vector extensions.
- KANs use learnable B-spline activations, differing from MLPs.
- Prior work characterized when KANs match or underperform MLPs in discovery accuracy.
- The paper asks if parameter efficiency survives deployment.
- The study uses identical trained weights for comparison.
- The paper is a cross-type announcement on arXiv.
Entities
Institutions
- arXiv