A Non-Pretrained Few-Shot Online Class Incremental Learning Design Based on Stochastic Adaptive Fourier Decomposition

Computer vision based few-shot class incremental learning (FSCIL) aims to learn new classes incrementally from very few labeled samples without forgetting previously learned classes. However, existing methods set some prerequisite assumptions, which are still far from realistic scenarios. In this paper, we focus on relaxing the various prerequisite assumptions as much as possible, i.e., non-pretrained fully online FSCIL without any data augmentation and exemplar selection. We achieve this learning by proposing a novel deep network, called SAFD_FSCIL, based on the mathematical theory - stochastic adaptive Fourier decomposition (SAFD). SAFD is a recently developed adaptive learning method based on a predefined dictionary. Its outstanding advantage is that it can learn the common features of multiple images at the same time, which enables our proposed SAFD backbone to be trained fully online, i.e., in just one epoch, without the need for pre-trained models, data augmentation and exemplar selection. In addition, we combine the learning ability of SAFD itself with the deep network. Each layer of the constructed SAFD backbone uses only one kernel for convolution, which can be calculated using two parameters. This greatly reduces the number of parameters, allowing us to use dynamic backbone duplication to mitigate catastrophic forgetting. Experiment results show that our proposed SAFD_FSCIL significantly outperforms the state-of-the-art results under the relaxed prerequisite assumptions.