KANs prove slower, costlier on embedded RISC-V than MLPs

A new research paper evaluates the deployment costs of Kolmogorov--Arnold Networks (KANs) compared to traditional Multilayer Perceptrons (MLPs) within hard-constrained recurrent physics-informed neural networks (HRPINNs) on embedded RISC-V systems. The study found that KANs, despite their parameter efficiency in training, were significantly slower and consumed more energy per integration step when deployed on a RISC-V RV64GC platform. Furthermore, KANs demonstrated lower dependability under INT8 quantization, with trajectories diverging earlier than comparable MLPs, suggesting MLPs remain a more robust default for embedded HRPINN applications. AI

IMPACT KANs show significant performance and energy drawbacks on embedded hardware, suggesting MLPs are a more practical choice for such applications.

RANK_REASON Research paper evaluating model performance on specific hardware. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• HRPINNs
• Kolmogorov--Arnold Networks
• RISC-V
• SiFive U74
• StarFive VisionFive 2

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →