Meta-Transfer Learning Through Hard Tasks.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Few-shot learning presents challenges due to limited labeled data.
• Deep neural networks (DNNs) often overfit with few samples, necessitating specialized meta-learning approaches.
• Existing meta-learning methods using pre-trained DNNs have limitations in direct weight transfer.

Purpose of the Study:

• To introduce Meta-Transfer Learning (MTL), a novel framework for few-shot learning.
• To enable effective weight transfer from pre-trained deep neural networks to new tasks.
• To improve learning efficiency and performance in few-shot classification.

Main Methods:

• Meta-transfer learning (MTL) learns to transfer DNN weights by optimizing scaling and shifting functions for each task.
• Introduced a hard task (HT) meta-batch scheme as a learning curriculum to enhance MTL efficiency.
• Conducted experiments on miniImageNet, tieredImageNet, and Fewshot-CIFAR100 (FC100) benchmarks.

Main Results:

• MTL combined with the HT meta-batch scheme achieved state-of-the-art performance on few-shot classification tasks.
• Demonstrated effectiveness in both supervised and semi-supervised settings.
• Ablation studies confirmed the contribution of both MTL and the HT scheme to rapid convergence and high accuracy.

Conclusions:

• Meta-transfer learning offers a powerful approach for few-shot learning by leveraging pre-trained deep networks.
• The hard task meta-batch scheme significantly boosts learning efficiency and classification accuracy.
• The proposed MTL framework represents a significant advancement in few-shot learning research.