Sampling in low confidence area of a targeted model
Résumé
Generative adversarial networks are traditionally used to generate data for itself (super resolution, procedural texture) or for helping training tasks (semi supervised training, zero shot learning, adversarial loss). In this paper, we aim to generate testing data. At first glance, this may seem useless because these data will require human annotation to ensure their relevancy and correct classes. But this human annotation is the main lock in many domains where there is a profusion of data (remote sensing, youtube video ...). Yet, the main idea of the offered framework is to sample especially in the low confidence area of a targeted model. This way, the human annotation cost is focused on hard samples. Despite this framework does not allow to mitigate the issue of false output with high confidence, it may be an interesting to tackle the rare event probability problem of estimating error probability in low confidence area of a targeted network.
Origine : Fichiers produits par l'(les) auteur(s)
Loading...