The paper introduces Meta-SGD, a new meta-learner designed for few-shot learning. Meta-SGD is an SGD-like meta-learner that can initialize and adapt any differentiable learner in just one step, making it highly efficient and easy to train. Unlike popular meta-learners like LSTM and MAML, Meta-SGD learns not only the learner initialization but also the update direction and learning rate, all in a single meta-learning process. This approach results in a meta-learner with higher capacity and better performance on few-shot learning tasks, including regression, classification, and reinforcement learning. The paper compares Meta-SGD with other meta-learners and demonstrates its superior performance through various experiments, showing that it can learn quickly from a few examples and adapt to new tasks effectively. The authors also discuss future directions, such as large-scale meta-learning and improving the generalization capacity of meta-learners.The paper introduces Meta-SGD, a new meta-learner designed for few-shot learning. Meta-SGD is an SGD-like meta-learner that can initialize and adapt any differentiable learner in just one step, making it highly efficient and easy to train. Unlike popular meta-learners like LSTM and MAML, Meta-SGD learns not only the learner initialization but also the update direction and learning rate, all in a single meta-learning process. This approach results in a meta-learner with higher capacity and better performance on few-shot learning tasks, including regression, classification, and reinforcement learning. The paper compares Meta-SGD with other meta-learners and demonstrates its superior performance through various experiments, showing that it can learn quickly from a few examples and adapt to new tasks effectively. The authors also discuss future directions, such as large-scale meta-learning and improving the generalization capacity of meta-learners.