Speaker: Sergio Márquez Carrero.

Abstract: Modern Deep Neural Networks (DNN) have significantly outperformed those employed over a decade ago in terms of accuracy. Nonetheless, the outputs generated by these models are poorly calibrated, causing substantial issues in a variety of decision-making applications. This situation has promoted the development of different techniques that aim to solve the calibration problem. However, previous work states that these methods present some limitations. Matrix Scaling is one of the techniques with a better performance, being able to calibrate complex distributions. Nevertheless, it fails in high-dimensional multiclass scenarios, where the calibrator overfits to the training samples as the number of parameters increases quadratically. For this reason, we aim to overcome this limitation by applying different data augmentation approaches on the logit distributions throughout the training phase of the calibrator. Thus, we explore the application of Mixup since, despite its simplicity, it improves the performance of the models in other tasks. However, it has been recently demonstrated that this strategy does not necessarily enhance calibration since it does not consider the uncertainty of the input distribution. This fact encourages the analysis of more complex data augmentation techniques, such as the use of generative models like Normalizing Flows to better estimate the logit distributions.