Few-Shot Learning With Class Imbalance

Few-Shot Learning (FSL) algorithms are commonly trained through Meta-Learning (ML), which exposes models to batches of tasks sampled from a meta-dataset mimic seen during evaluation. However, the standard training procedures overlook real-world dynamics where classes occur at different frequencies. While it is generally understood that class imbalance harms performance supervised methods, limited research examines impact on FSL evaluation task. Our analysis compares 10 state-of-the-art meta-learning and methods distributions rebalancing techniques. results reveal 1) some display natural disposition against while most other approaches produce drop by up 17% compared balanced task without appropriate mitigation; 2) many will not automatically learn balance exposure imbalanced tasks; 3) classical strategies, such as random oversampling, can still be very effective, leading performances should overlooked.