Latent attribute augmented network for few-shot class-incremental learning

Abstract

Few-Shot Class-Incremental Learning (FSCIL) aims to learn knowledge about new classes continually with limited labeled data, where preserving knowledge about old classes and fast adaptation to new classes play important roles in FSCIL models. Traditional approaches usually learn discriminative global features among base classes and the feature extraction model is directly applied to novel categories. However, the global features, though effective for base class discrimination, are less generalized to novel categories. Therefore, we focus on the more generalizable local parts and propose the Latent Attribute Augmented Network (LAAN) to enhance the discriminative local features. Specifically, we automatically learn some latent attributes that are shared among different classes and utilize them to focus on the key local regions of the images. Furthermore, we construct a transformer-based knowledge interaction module, to fuse the information of latent attributes and the local features, which obtains more discriminative features to classify different classes. Our method has achieved superior performance across three benchmark datasets: CIFAR100, mini-ImageNet, and CUB200.