Neural network representations proved to be relevant for many computer vision tasks such as image classification, object detection, segmentation or instance-level image retrieval. A network is trained for one particular task and requires a large number of labeled data. We propose in this thesis solutions to extract the most information with the least supervision. First focusing on the classification task, we examine the active learning process in the context of deep learning and show that combining it to semi-supervised and unsupervised techniques boost greatly results. We then investigate the image retrieval task, and in particular we exploit the spatial localization information available ``for free'' in CNN feature maps. We first propose ...