Abstract
Machine learning is becoming an increasingly popular method in science, as it commonly outperforms traditional methods for pattern recognition and prediction of new data. At the same time, such methods are commonly criticised for hindering understanding of the phenomena, such that all we are left with are predictions without explanations. In this talk, I challenge the notion that ML itself is the sole cause of these issues, arguing that the root of the problem is often the datasets used, rather than the analysis method per se.
I will look at some concrete objections to ML methods regarding the reliability and interpretability of the models produced, and respond to these by suggesting some solutions based on more traditional approaches and assumptions in science. Specifically, I argue that data-driven ML models are comparable with phenomenological models, but that controlling and intervening on the input variables is needed to learn a causal model of the system, as has always been the case. I will also present some suggested methods for testing this in practice. Overall, regardless of its novelty in some respects, I argue that machine learning is in many ways continuous with traditional methods in science, and that its problems are to a large extent solvable using established methods.