Coarse-To-Fine Incremental Few-Shot Learning

Abstract

Different from fine-tuning models pre-trained on a large-scale dataset of preset classes, class-incremental learning (CIL) aims to recognize novel classes over time without forgetting pre-trained classes. However, a given model will be challenged by test images with finer-grained classes, e.g., a basenji is at most recognized as a dog. Such images form a new training set (i.e., support set) so that the incremental model is hoped to recognize a basenji (i.e., query) as a basenji next time. This paper formulates such a hybrid natural problem of coarse-to-fine fewshot (C2FS) recognition as a CIL problem named C2FSCIL, and proposes a simple, effective, and theoretically-sound strategy Knowe: to learn, normalize, and freeze a classifier's weights from fine labels, once learning an embedding space contrastively from coarse labels. Besides, as CIL aims at a stability-plasticity balance, new overall performance metrics are proposed. In that sense, on CIFAR-100, BREEDS, and tieredImageNet, Knowe outperforms all recent relevant CIL/FSCIL methods that are tailored to the new problem setting for the first time.Preprint. Under review.

Research on the Fusion of Artificial Intelligence and 6G in Computer Network Technology

“…Coarse and Fine Learning Coarse and fine learning has been an important topic in computer vision and machine learning. Numerous methods [2,7,8,26,27,29,35,40,42] and theoretical studies [10] have been proposed to address this problem, with the goal of leveraging coarse-grained labeled data to improve fine-grained recognition. On the one hand, several methods have been proposed to tackle the coarse and fine learning problem.…”

“…For instance, Stretcu et al [26] proposed a coarse-to-fine curriculum learning method that can dynamically generate training sample sequences based on task difficulty and data distribution, thereby accelerating model convergence and improving generalization ability. Xiang et al [40] proposed a coarse-to-fine incremental few-shot learning method that can use coarse-grained labels to perform contrastive learning on the embedding space and then used fine-grained labels to normalize and freeze the classifier weights, thereby solving the class incremental problem. Sun et al [27] developed a dynamic metric learning method that can adaptively adjust the metric space according to different semantic scales, thereby improving the performance of multi-label classification and retrieval.…”