New SDZE framework enables training of 10M-dimensional PINNs on single GPU

Researchers have developed a new framework called the Stochastic Dimension-free Zeroth-order Estimator (SDZE) to address memory and computational constraints in training physics-informed neural networks (PINNs). SDZE achieves dimension-independent complexity for both spatial and memory requirements by synchronizing random numbers across perturbations to mitigate variance explosion. This method also reduces parameter exploration variance, enabling the training of PINNs with up to 10 million dimensions on a single NVIDIA A100 GPU, offering significant improvements in efficiency. AI

IMPACT This new method could significantly reduce the computational resources required for training complex physics-informed neural networks, potentially accelerating research in fields relying on solving high-dimensional partial differential equations.

RANK_REASON Academic paper detailing a new method for training neural networks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Common Random Numbers Synchronization
• NVIDIA A100 GPU
• partial differential equations
• physics-informed neural networks
• SDZE
• Stochastic Dimension Zeroth-Order Estimator
• Zhangyong Liang

• paper
• infra

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