Abstract:Few-shot class-incremental learning (FSCIL) aims to continually fit new classes with limited training data, while maintaining the performance of previously learned classes. The main challenges are overfitting the rare new training samples and forgetting old classes. While catastrophic forgetting has been extensively studied, the overfitting problem has attracted less attention in FSCIL. To tackle overfitting challenge, we design a new ensemble model framework cooperated with data augmentation to boost generalization. In this way, the enhanced model works as a library storing abundant features to guarantee fast adaptation to downstream tasks. Specifically, the multi-input multi-output ensemble structure is applied with a spatial-aware data augmentation strategy, aiming at diversifying the feature extractor and alleviating overfitting in incremental sessions. Moreover, self-supervised learning is also integrated to further improve the model generalization. Comprehensive experimental results show that the proposed method can indeed mitigate the overfitting problem in FSCIL, and outperform the state-of-the-art methods.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to effectively deal with the problems of over - fitting and catastrophic forgetting in Few - Shot Class - Incremental Learning (FSCIL). Specifically, the goal of FSCIL is to continuously adapt to new classes with limited training data of new classes while maintaining the performance on previously learned classes. The main challenges are: 1. **Over - fitting**: Since the training samples of new classes are very limited, the model is prone to quickly memorize these limited samples, resulting in poor generalization ability on new tasks. 2. **Catastrophic forgetting**: When learning new classes, the model may forget the old classes that have been learned before, which is a common problem in incremental learning. To address these challenges, the paper proposes a new integrated model framework and combines data augmentation and self - supervised learning strategies to improve the generalization ability of the model and reduce over - fitting. Specific methods include: 1. **Multi - input and multi - output integrated model**: By using the multi - input and multi - output integrated model structure, the model can provide diverse feature templates, so as to better adapt to downstream tasks. 2. **Spatially - aware data augmentation**: A background - level data augmentation method is designed to introduce diversity while protecting the main part of the sample, thereby alleviating the over - fitting problem. 3. **Self - supervised learning**: Integrated self - supervised learning with mixed - feature compatibility makes the model pay more attention to general and universal representations, further enhancing the generalization ability of the model. Through these methods, the paper aims to alleviate the over - fitting problem in FSCIL and has verified its effectiveness on multiple benchmark datasets. The results show that this method is significantly superior to the existing state - of - the - art methods in performance.

Enhanced Few-Shot Class-Incremental Learning via Ensemble Models

A Cognition-Driven Framework for Few-Shot Class-Incremental Learning

Few-shot class-incremental learning based on representation enhancement

A survey on few-shot class-incremental learning

Self-supervised Contrastive Feature Refinement for Few-Shot Class-Incremental Learning.

Few-Shot Incremental Learning with Continually Evolved Classifiers

Model Attention Expansion for Few-Shot Class-Incremental Learning

Pseudo Initialization Based Few-Shot Class Incremental Learning

Few-Shot Class-Incremental Learning Via Class-Aware Bilateral Distillation

TLCE: Transfer-Learning Based Classifier Ensembles for Few-Shot Class-Incremental Learning

FSCIL-EACA: Few-Shot Class-Incremental Learning Network Based on Embedding Augmentation and Classifier Adaptation for Image Classification

Few Shot Class Incremental Learning via Efficient Prototype Replay and Calibration

S3C: Self-Supervised Stochastic Classifiers for Few-Shot Class-Incremental Learning

Knowledge Representation by Generic Models for Few-Shot Class-Incremental Learning

Bias Mitigating Few-Shot Class-Incremental Learning

A streamlined Approach to Multimodal Few-Shot Class Incremental Learning for Fine-Grained Datasets

Jointly Optimized Classifiers for Few-Shot Class-Incremental Learning

Rethinking Few-shot Class-incremental Learning: Learning from Yourself

Synthesized Feature based Few-Shot Class-Incremental Learning on a Mixture of Subspaces

Analogical Learning-Based Few-Shot Class-Incremental Learning