Speaker: Sergio Álvarez Balanya.
Abstract: Calibration is a desirable property of pattern recognition systems, especially when their predictions are going to be used to make decisions. In our group, we are used to dealing with calibration in classification tasks such as speaker recognition, glass comparison, and benchmark datasets—e.g. CIFAR-10. However, the theoretic framework is general to any machine learning task, thus not limited to classification. In this seminar, we show how to obtain in regression problems two common diagnostic tools for calibration: reliability plots and an ECE like error measure. Based on http://proceedings.mlr.press/v80/kuleshov18a/kuleshov18a.pdf.