This paper proposes a novel meta-learning approach called meta-transfer learning (MTL) for few-shot learning. MTL leverages pre-trained deep neural networks (DNNs) to adapt to new tasks with limited labeled data. The key idea is to use scaling and shifting operations on DNN weights to transfer knowledge from large-scale data to few-shot tasks. Additionally, the paper introduces a hard task (HT) meta-batch strategy to improve learning efficiency by focusing on challenging tasks. The MTL method consists of three phases: first, training a DNN on large-scale data; second, learning scaling and shifting parameters for the DNN neurons; and third, testing the model on unseen tasks. The HT meta-batch strategy is designed to select and re-sample hard tasks based on previous failures, enabling faster convergence and better performance. Experiments on two benchmarks, miniImageNet and Fewshot-CIFAR100, show that MTL outperforms existing methods in few-shot learning tasks. The proposed method achieves high accuracy and fast convergence, especially when using the HT meta-batch strategy. The results demonstrate that MTL is effective in adapting to new tasks with limited data, and that the HT meta-batch strategy significantly improves learning performance. The paper also shows that the MTL method is more efficient than traditional meta-learning approaches, as it uses fewer tasks and parameters while achieving better results.This paper proposes a novel meta-learning approach called meta-transfer learning (MTL) for few-shot learning. MTL leverages pre-trained deep neural networks (DNNs) to adapt to new tasks with limited labeled data. The key idea is to use scaling and shifting operations on DNN weights to transfer knowledge from large-scale data to few-shot tasks. Additionally, the paper introduces a hard task (HT) meta-batch strategy to improve learning efficiency by focusing on challenging tasks. The MTL method consists of three phases: first, training a DNN on large-scale data; second, learning scaling and shifting parameters for the DNN neurons; and third, testing the model on unseen tasks. The HT meta-batch strategy is designed to select and re-sample hard tasks based on previous failures, enabling faster convergence and better performance. Experiments on two benchmarks, miniImageNet and Fewshot-CIFAR100, show that MTL outperforms existing methods in few-shot learning tasks. The proposed method achieves high accuracy and fast convergence, especially when using the HT meta-batch strategy. The results demonstrate that MTL is effective in adapting to new tasks with limited data, and that the HT meta-batch strategy significantly improves learning performance. The paper also shows that the MTL method is more efficient than traditional meta-learning approaches, as it uses fewer tasks and parameters while achieving better results.