The Falsificationist View of Machine Learning

Abstract

Machine learning pushes the frontiers of algorithmic achievements, though the striving for state-of-the-art performance often obscures the fragility of enforcing decisions amid uncertainty. This paper interprets machine learning within Karl Popper’s epistemology. We assess machine learning paradigms’ fit for falsificationism and argue that the new interpretation can improve robustness. Though the price is to accept unambiguous decisions, the restriction of the hypothesis space still adds value. The context for our work is established by comparison with similar techniques and highlighting its limitations.