Are We Underestimating Overfitting?

It is a well-established dogma in quantitative structure-activity relationships (QSAR) that parsimonious models generalize best and that overfitting must be avoided. As the number of fitted parameters in a model approaches the number of training examples, the training errors decrease and the external test set errors usually increase substantially. This makes intuitive sense, but recent publications on overfitting and overparameterization, and the dramatic rise in complex but very useful deep learning models, have suggested that formally overparameterized machine learning models may recover their ability to predict external data accurately. This counterintuitive idea is supported by several information theoretic arguments and by modeling of potentially overfitted synthetic and real data. The implication is that supernumerary model parameters contain additional information on SAR that may contribute to model predictive accuracy for unseen data. Here, we discuss the implications of this understanding of overfitting and overparameterization, discuss its implications for QSAR and quantitative structure-property relationship (QSPR) modeling, and provide illustrative examples of the ability of overfitted ML models to predict test set data well. "It is better to be approximately right than precisely wrong" - Warren Buffett.