Analyse et méthodes de recommendation pour l'apprentissage incrémental par classe

Continual learning is an essential aspect of human intelligence, while it is a major challenge for artificial learning agents. In artificial intelligence, continual learning addresses the problem of learning adaptively from a stream of data to develop more complex skills and knowledge. In this thesis, we focus on Class-Incremental Learning (CIL), which involves training a deep neural network sequentially to recognize an increasing number of classes, without access to previous data. On the one hand, CIL is challenging due to catastrophic forgetting, i.e. the tendency of deep models to forget previously acquired information when confronted with new information. On the other hand, CIL is attractive for its reduced storage and computational requirements compared to re-training the model from scratch every time an update is needed. Our work has two main axes. First, we analyze the role of pre-training in the performance of CIL algorithms. We present a statistical analysis of the main factors influencing incremental learning performance. We also conduct an experimental study that highlights the importance of training the initial model without overfitting on the initial classification task. Second, we tackle the problem of recommending a CIL algorithm adapted to a given use case. In particular, we recommend an adequate pair of a neural architecture and a CIL algorithm, under the constraint of a given number of parameters for the model at inference. In addition, we propose a stream simulation method to assess the performance of CIL algorithms and recommend an adequate algorithm for a given use case.We hope this work will contribute to making deep learning more frugal in terms of memory and computational resources.