Un séminaire de l'équipe A3SI du LIGM (unité mixte de recherche de l'Université Paris Est) aura lieu le mercredi 26 avril à 13h45, dans la salle de réunion B 412 du groupe IMAGINE (ENPC - Bat. Coriolis).
Abstract: My talk is is going to be focused on semi-supervised training of structured-output neural networks. The problem is of broad interest, because training deep neural networks requires large amounts of annotated data, and producing structured annotations is costly. On the other hand, unannotated input data is often cheap to acquire. I will show that this unannotated data can be used for training structured-output models in a semi-supervised scenario. The idea is inspired by the Generative Adversarial Networks, and consists in generating an error signal for the unlabelled data by means of a discriminator neural network. To be source of a useful error signal, the discriminator needs to capture quality of the structured output. This can be achieved by training the discriminator to differentiate between structured outputs obtained for the labelled training data (qualitatively better) and the outputs obtained for unlabelled data, not used for training (qualitatively worse). Initial experiments suggest that this approach may lead to decreasing the volume of labels required to attain a given precision by a significant factor.