Balancing Old and New Classes for Class-Incremental Learning Methods

In the Class-Incremental Learning setting, instead of being trained on the complete dataset at once, machine learning models are trained sequentially on new classes while training data from previously learned classes are no longer available.This training paradigm, closer to the human learning process, is challenging for deep neural network models due to catastrophic forgetting: when learning new classes, the model tends to forget the old classes resulting in a significant deterioration of its performance.This thesis focuses on Class-Incremental Learning for deep neural networks trained for image classification.In order to mitigate catastrophic forgetting, we explore the important trade-off between old and new classes and propose three distinct approaches meant to be seamlessly combined with state-of-the-art approaches for Class-Incremental Learning.First, we propose to balance the contribution of samples from old and new classes in the classification loss by introducing the Balanced Softmax Cross Entropy for Class-Incremental Learning with a rehearsal memory and then extend it to the more challenging setting of Exemplar-Free Class-Incremental Learning where it is impossible to save any samples of past classes.Secondly, we propose to balance the training dataset itself by compensating for the limited size of the replay memory by using additional synthetic samples of past classes generated using pre-trained text-to-image diffusion models.Finally, we propose to pre-train the feature extractor of models designed for Exemplar-Free Class-Incremental Learning better by leveraging knowledge about possible future classes during the initial training step.To do so, our method relies on pre-trained text-to-image diffusion models to generate synthetic samples of future classes in a fast and economical manner.Extensive experiments on challenging benchmarks for Class-Incremental Learning highlight that our proposed methods significantly improve the performances of state-of-the-art approaches.