Symmetries in PAC-Bayesian Learning
Researchers have developed new theoretical guarantees for machine learning models that utilize symmetries, extending beyond compact groups and invariant data distributions. The study adapts and tightens existing bounds within the PAC-Bayes framework, demonstrating its applicability to various PAC-Bayes bounds. Experiments on datasets with non-uniform and non-compact transformations validate the theory, showing improved results and providing evidence for the preference of symmetric models with symmetric data. AI
IMPACT Provides theoretical grounding for using symmetries in ML models, potentially leading to more robust and efficient algorithms.