TA-MTL: Task-Adaptive Meta Transfer Learning for Few-shot Learning

Meta-learning-based methods have been widely adopted to address few-shot learning challenges. After training on a diverse set of tasks, meta-learning models can achieve strong generalization performance on new tasks. However, meta-learning still faces several critical limitations: (1) During meta-training, the base learner is overfitted to the training tasks, leading to insufficient generalization capability on new tasks; (2) Deep neural networks are prone to overfitting when only a small number of labeled samples are available, meta learning therefore often uses shallow networks, which limits its performance on complex tasks. To address these issues, we propose a task-adaptive meta transfer learning (TA-MTL) to alleviate the problem of "catastrophic forgetting" and solve the problem of overfitting. The meta transfer learning optimizes and improves the model's new-task adaptability through contrastive learning, and the task-adaptive algorithm can learn an unbiased base learner and prevent it from over-performing in meta training tasks. TA-MTL learns an unbiased base learner and prevents overfitting to meta-training tasks. Experiments on benchmarks (Mini-ImageNet and FC100) demonstrate that TA-MTL exhibits satisfactory performance in few-shot Learning.