Two new research papers explore the application of neural networks and transfer learning to high-dimensional Ising models. The first paper investigates out-of-distribution performance of various neural architectures, finding that Transformer-based models and convolutional neural networks employ different statistical strategies that can lead to apparent robustness without true physical rule learning. The second paper introduces Trans-Ising, a transfer learning method designed to improve Ising model estimation by effectively utilizing auxiliary datasets, demonstrating lower estimation errors compared to target-only methods. AI
IMPACT These studies highlight potential pitfalls in applying neural networks to scientific discovery and propose methods to improve model performance using transfer learning.
RANK_REASON Two new arXiv papers on machine learning applications to statistical physics models.
AI-generated summary · Google Gemini · from 3 sources. How we write summaries →
