Event Details

SuperNNova: Preparing the next decade SN samples with Bayesian Neural Networks

Abstract:

With future surveys such as LSST discovering up to hundred thousand SNe, there is not nearly enough telescope time to spectroscopically confirm each supernova type. Photometric classification methods provide a typing using only the evolution of flux with time, light-curve, for a given supernova. In this talk, I will present SuperNNova, an open source Neural Network framework that is able to obtain photometric samples with purities similar to those of spectroscopically selected samples, <2% contamination, as well to provide a Bayesian interpretation. I will discuss the impact of photometric classifiers in dark energy studies and selection of promising candidates for follow-up, and often neglected pitfalls of machine learning algorithms. I will emphasize the importance of properly calibrated classification probabilities and the additional information that provide epistemic uncertainties when dealing with representativeness issues and out-of-distribution events.