Speaker: Sergio Márquez
Abstract: Today’s Deep Neural Networks (DNNs) have achieved high performance in accuracy, far exceeding the ones used ten years ago. Nevertheless, the outputs provided by these modern networks are less well calibrated, becoming a major problem in many applications. We evaluate the performance of various calibration methods on outputs provided by a neural network (EfficientNet-B0) trained with image datasets (CIFAR-3, CIFAR-10, CIFAR-100). We propose a model based in the application of Normalizing Flows that improves calibration and works for certain cases for which methods such as Temperature Scaling (TS) are not very effective. Experiments carried out show that the method implemented with a NICE flow (nonlinear independent component estimation) produce similar results to the one using TS in terms of calibration on CIFAR-3, besides, it works better in a case where the classes are rotated.