Is Occam’s razor losing its edge? New perspectives on the principle of model parsimony

Parsimony has long served as a criterion for selecting between scientific theories, hypotheses, and models. Yet recent years have seen an explosion of incredibly complex models, such as deep neural networks (e.g., for 3D protein folding) and multi-model ensembles (e.g., for climate forecasting). This perspective aims to re-examine the principle of model parsimony in light of the recent advances in science and technology. We review recent developments such as the discovery of double descent of prediction error, the increasing appreciation of the context-sensitivity of data and misspecification of scientific models, as well as the new types of models and modeling tools available to scientists. We integrate these results to reevaluate the utility of the parsimony principle as a proxy for desirable model traits, such as predictive accuracy, interpretability, utility in guiding future research, resource efficiency, and others. We highlight the need for a nuanced, context-dependent application of the parsimony principle, acknowledging situations where more complex models may be more appropriate or even necessary for achieving scientific goals.