New research explores finite-difference methods for PINNs

A new paper explores the use of finite-difference (FD) methods for computing derivatives in Physics-Informed Neural Networks (PINNs), presenting it as an alternative to automatic differentiation (AD). The research demonstrates that FD methods, when properly calibrated, can match AD in accuracy while being faster and more memory-efficient across various batch sizes. The paper also highlights potential inaccuracies in standard PyTorch autograd for certain neural architectures and suggests FD as a more feasible approximation for per-sample derivatives at relevant scales. AI

IMPACT This research could lead to more efficient and accurate training of physics-informed neural networks, potentially impacting scientific machine learning applications.

RANK_REASON The cluster contains an academic paper detailing a new method for derivative computation in neural networks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• automatic differentiation
• batch normalization
• Finite differences for the convection-diffusion equation
• Maciej J. Mikulski
• physics-informed neural networks
• PyTorch

• paper
• infra

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