[samovar-armedia] séminaire ARMEDIA jeudi 7 novembre: Deep Active Learning with Rationales

L’équipe ARMEDIA organise un séminaire autour de l’IA le jeudi 7 novembre à 14h en salle H218 (bâtiment Etoile, campus Evry).

Le thème de la présentation et un résumé en anglais sont disponibles ci-dessous:

Deep Active Learning with Rationales
By Paul Guelorget (Ph D student with ARMEDIA team)

Active learning techniques aim to solve the lack of reliable groundtruth information in machine learning problems, by exploiting the knowledge of a human labeler called oracle. During each active learning iteration, documents whose labels are the most needed are labeled by the oracle before they are added to the training set. Beyond a simple label, it is possible to exploit expert knowledge even further by asking what parts of the document are the most responsible for class membership. We propose to use such rationales to improve the learning curve of a fully-convolutional neural network dedicated to text classification, by constraining Class Activation Maps (CAM) extracted from the network itself. It appears that the raw transfer of detailed spatial knowledge drastically improves the network’s performances at the earliest stages of active learning. Also, the injection of scarcer spatial information to simulate human-like rationales appears to be beneificial for the learning curve.