On Few-Annotation Learning and Non-Linearity in Deep Neural Networks
Date:
Learning something new in real life does not necessarily mean going through a lot of examples in order to capture the essence of it. Humans are able to build upon prior experience and have the ability to adapt, allowing to combine previous observations with only little evidence for fast learning. This is particularly the case for recognition tasks, for which we are often capable of differentiating between two distinct objects after having seen only a few examples of them. In this talk, I will develop three different contributions for Machine Learning with limited labels, and more specifically for Computer Vision tasks, addressing theoretical, algorithmic and experimental aspects. In a first contribution, we are interested in bridging the gap between theory and practice for popular Meta-Learning algorithms used in Few-Shot Classification. We make connections to Multi-Task Representation Learning, which benefits from solid theoretical foundations, to verify the best conditions for a more efficient meta-learning. Then, to leverage unlabeled data when training object detectors based on the Transformer architecture, we propose an unsupervised pretraining approach that improves contrastive learning for object detectors through the introduction of the localization information. Finally, we present the first theoretically sound tool to track non-linearity propagation in deep neural networks, with a specific focus on computer vision applications. Our proposed affinity score allows us to gain insights into the inner workings of a wide range of different architectures and learning paradigms. We present extensive experimental results that highlight the practical utility of the proposed affinity score and its potential for long-reaching applications.