Robust validation steps for clip classification

Full Title: “Robust validation steps for clip classification”

Authors: Mariano Rodríguez, Liam Schoneveld and Vincent Garcia

Abstract: The classic approach to clip classification consist of computing scores per class and identifying the most frequent top-score classes. Usually, predictions are always reported; and consequently, we are either right or wrong. However, in most applications, erroneous predictions will carry substantial negative effects. Ideally, we should communicate only correct predictions and abstain from reporting erroneous ones. Two novel methods are shown to come near this goal. Their main advantages are: on the one hand, most false positives are successfully identified; on the other hand, the best levels of accuracy are maintained, if not improved.