There is no data-generating distribution (www.argmin.net)

In a thought-provoking final lecture of his graduate course "Patterns, Predictions, and Actions," a professor shared insights on the fundamental misconceptions in machine learning, particularly the reliance on the concept of a data-generating distribution. He argued that this notion is a myth, asserting that machine learning doesn’t require a predefined probability distribution to operate effectively. Instead, he emphasized focusing on how models are constructed based on sample populations and the assumption frameworks that guide decision-making processes, such as batch learning, online learning, and empiricist learning. This perspective challenges long-standing assumptions within the AI/ML community, suggesting that traditional teaching methods may hinder the understanding of machine learning’s true nature. By promoting a "distribution-free" approach, he encourages practitioners to recognize that randomness in data may be an artificial construct, not a core component of machine learning methodology. This shift could have significant implications for both the development of machine learning models and their application in real-world decision-making, urging engineers to rethink the necessity of statistical models in an increasingly data-driven culture.