Prediction with Uncertainty
vendredi 17 janvier 2020, 13h30 - 14h30
Uncertainty estimation in neural network predictions is not yet well studied (except in the recent and controversial field called Bayesian Deep Learning). Moreover, in the academic and business worlds, some misinterpretations persist about the probabilities as outputs of a supervised classification: a probability of belonging to a category or its associated vector across all categories does not fully express information on uncertainty. Indeed, most practitioners extract the maximum entry of a probability vector, which is hacky and not sufficient for a more thorough interpretation of uncertainty. At the same time, in many sensitive applications where security, health or even justice are involved, it seems that this uncertainty estimation is crucial. At the crossroads of Bayesian statistics and frequentist neural networks, we tackle uncertainty in Supervised Machine Learning in an engineering but yet principled point of view.