Yara M. Bahram

Machine learning (ML) relies heavily on regularization, as it allows better generalization to unknown data, even with imperfect optimization procedures and datasets. There are, however, major problems with regularization that have surprisingly received little attention so far. Regularization methods traditionally avoid overfitting. But when overfitting happens, they usually fail to fight it and cannot bring the machine learning model out of its adverse situation. As a result, the stage of overfitting is thought of as a dead-end in the ML community. This idea has consequently forced the system designers to perform expensive hyperparameter searches, retraining the model from scratch every time with new configurations. It has also led to similar complications in the ever-changing dynamic usages of static pre-trained models. Defying the overfitting dead-end misconception, we argue that even a model that is overfitting includes useful information about the task at hand, and being able to adjust the regularization strength by using this information after the overfitting further solidifies this suggestion. This process ideally requires the disentanglement of the regularization process from the initial stages of the training phase and being able to apply the regularization as a post-processing step with low cost. This would allow adjusting the regularization strength of pre-trained models efficiently, which is an increasingly viable concept given the prevalent use of large neural networks today in many domains and applications. There already are implicit and explicit traces of this idea, which we call Post-Regularization, in a wide range of existing works from several domains. However, there exists no unified view of this concept. In this work, we formalize Post-Regularization, and provide a novel taxonomy of regularization, from the perspective of when the regularization is applied with respect to the model training to help bring together the ideas that can potentially be further explored in this area. We hope that this work attracts more attention to Post-Regularization and provides a foundation for future related work.

A wide variety of fairness metrics and eXplainable Artificial Intelligence (XAI) approaches have been proposed in the literature to identify bias in machine learning (ML) models that are used in critical real-life contexts. In this paper, we present an approach that ties the evaluation of the fairness of machine learning models to the expense (or conversely the simplicity) of generating counterfactual explanations (CFE). Considering the relative importance of each feature class towards obtaining the desired outcome, we attribute different scores to each feature value and propose model-agnostic concepts of Qualification and Effort, respectively, indicating the candidate's strength and the adjustments that must be made to the candidate's profile in order to alter the classifier's output. To ensure a fair comparison, we further group candidates based on their Qualification scores. We additionally utilize KL-divergence to capture the complex disparities between distributions of the CFE Efforts among the sensitive groups. We empirically evaluate fairness among different sensitive classes for several classifiers on two real-world datasets. We find that the majority of classifiers exhibit unfair propensities towards the minority groups. However, some models exacerbate the bias more than others. Also, we find that classifiers tend to behave differently for members of each Qualification group. Our results show that our method can successfully be used for fairness evaluation of classification models based on CFEs.