Boosting meta-training with base class information for robust few-shot learning

Abstract

Few-shot learning aims to develop classifiers that can recognize new, unseen classes with only a few labeled examples. Meta-learning, with methods like Model-Agnostic Meta-Learning (MAML) and Prototypical Networks, has emerged as a key approach for this challenge. A recent advancement, Meta-Baseline, utilizes sequential pre-training and meta-training to achieve state-of-the-art performance. However, during meta-training, designed to further adapt the model to few-shot classification tasks, the class transferability gained during pre-training can be compromised, leading to suboptimal performance. We propose an end-to-end training paradigm with two alternating loops. The outer loop calculates cross-entropy loss across the entire training set while updating only the final linear layer. The inner loop uses the original meta-learning method to calculate the loss and integrate gradients from the outer loss to guide parameter updates. This approach enables effective adaptation to few-shot learning tasks while preserving robust generalization, outperforming existing baselines. Our findings suggest that leveraging information from the entire training set and the meta-learning training paradigm could mutually enhance one another. Our extensive experiments on the $mini$ImageNet, $tiered$ImageNet, and CUB datasets demonstrate the effectiveness of our method. In the 5-way 1-shot setting, we achieve accuracies of 64.01%, 69.73%, and 68.07%, outperforming the best baselines by 0.94%, 1.11%, and 1.94%, respectively. In the 5-way 5-shot setting, our accuracies are 81.00%, 84.91%, and 83.44%, outperforming the best baselines by 1.74%, 1.17%, and 3.37%, respectively. Additionally, our framework is model-agnostic, yielding significant performance improvements when integrated with existing baseline training frameworks, providing an approximate 1% performance boost.