Machine learning in physics: the pitfalls of poisoned training sets. Issue 4 (11th September 2020)

• Record Type: Journal Article
• Authors: Fang, Chao; Barzeger, Amin; Katzgraber, Helmut G
• Is Part Of: Machine learning: science and technology. Volume 1:Issue 4(2020)
• Publication Date: 2020-09-11
• Subjects: poisoned training sets -- spin glasses -- phase transitions -- condensed matter physics; 006.31
• DOI: 10.1088/2632-2153/aba821
• Languages: English
• ISSNs: 2632-2153

Abstract

Abstract: Known for their ability to identify hidden patterns in data, artificial neural networks are among the most powerful machine learning tools. Most notably, neural networks have played a central role in identifying states of matter and phase transitions across condensed matter physics. To date, most studies have focused on systems where different phases of matter and their phase transitions are known, and thus the performance of neural networks is well controlled. While neural networks present an exciting new tool to detect new phases of matter, here we demonstrate that when the training sets are poisoned (i.e. poor training data or mislabeled data) it is easy for neural networks to make misleading predictions.