Machine learning in physics: The pitfalls of poisoned training sets

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.

arXiv comment5 pages, 3 figures, 3 tables

References in corpus (17)

• TensorFlow: A system for large-scale machine learning
• Exchange Monte Carlo Method and Application to Spin Glass Simulations
• Machine learning phases of matter
• Machine Learning Phases of Strongly Correlated Fermions
• Universality in three-dimensional Ising spin glasses: A Monte Carlo study
• Detection of phase transition via convolutional neural network
• Manipulating Machine Learning: Poisoning Attacks and Countermeasures for Regression Learning
• On the Phase Structure of the 3D Edwards Anderson Spin Glass
• Nature of the spin-glass phase at experimental length scales
• Thermodynamic glass transition in a spin glass without time-reversal symmetry
• Dynamical transition in the D = 3 Edwards-Anderson spin glass in an external magnetic field
• Heavy tails in the distribution of time-to-solution for classical and quantum annealing
• Dynamical scaling in Ising and vector spin glasses
• Phase transition encoded in neural network
• Machine learning generated configurations in presence of a conserved quantity: a cautionary tale
• Learning to find order in disorder
• Spectrum Data Poisoning with Adversarial Deep Learning