New Meta-Learning Method Boosts Efficiency Without Labeled Data

Meta-learning without labeled data is crucial for real-world applications where data labeling is expensive or restricted. Existing Data-Free Meta-Learning (DFML) methods often rely on model inversion to generate synthetic training data, a process that is computationally intensive and challenging due to the need to accurately replicate original data distributions. This paper introduces a new meta-learning setting that bypasses model inversion entirely. Instead, it jointly leverages pre-trained models and readily available unlabeled data. The core idea is to generate meta-training tasks by assigning "soft labels" to unlabeled data using the pre-trained models. To ensure the quality of these generated tasks, the method incorporates a task-weighting mechanism that considers both task confidence and class distribution balance. Extensive experiments demonstrate that this novel approach offers substantial benefits. It achieves up to a 104-fold speedup in computational cost and improves few-shot classification accuracy by 8.4% to 36.4% compared to state-of-the-art DFML methods. This makes meta-learning more accessible and efficient for scenarios where labeled data is scarce.