This paper investigates first-order meta-learning algorithms, focusing on their ability to quickly adapt to new tasks. The authors analyze a family of algorithms that learn an initialization for neural network parameters, which can be fine-tuned efficiently on new tasks using only first-order derivatives. This family includes first-order MAML (FOMAML), an approximation of MAML that ignores second-order derivatives, and Reptile, a new algorithm that iteratively samples tasks, trains on them, and updates the initialization towards the trained weights. The paper expands on previous results showing that first-order meta-learning algorithms perform well on few-shot classification benchmarks. It provides theoretical analysis to understand why these algorithms work, showing that they optimize for within-task generalization. The authors also introduce Reptile, which is closely related to FOMAML and is simpler to implement. Unlike FOMAML, Reptile does not require a training-test split for each task, making it more natural in certain settings. The paper evaluates Reptile on the Mini-ImageNet and Omniglot datasets, showing that it performs slightly better than FOMAML on Mini-ImageNet and slightly worse on Omniglot. It also demonstrates that Reptile converges to a solution that is different from the minimizer of the expected loss. Theoretical analysis shows that Reptile's performance is driven by maximizing the inner product between gradients from different minibatches on the same task, which improves generalization. The paper also explores the effectiveness of Reptile in different settings, including the transductive setting where the model classifies the entire test set at once. It shows that Reptile performs well in this setting, suggesting that further research should focus on its use of batch normalization during testing. The authors also compare different inner-loop gradient combinations and find that Reptile benefits from taking many inner-loop steps, which is consistent with the optimal hyper-parameters found in previous experiments. Overall, the paper provides a comprehensive analysis of first-order meta-learning algorithms, showing that Reptile is a promising approach that can be implemented simply and effectively. The results suggest that Reptile is a strong baseline for meta-learning in various machine learning problems.This paper investigates first-order meta-learning algorithms, focusing on their ability to quickly adapt to new tasks. The authors analyze a family of algorithms that learn an initialization for neural network parameters, which can be fine-tuned efficiently on new tasks using only first-order derivatives. This family includes first-order MAML (FOMAML), an approximation of MAML that ignores second-order derivatives, and Reptile, a new algorithm that iteratively samples tasks, trains on them, and updates the initialization towards the trained weights. The paper expands on previous results showing that first-order meta-learning algorithms perform well on few-shot classification benchmarks. It provides theoretical analysis to understand why these algorithms work, showing that they optimize for within-task generalization. The authors also introduce Reptile, which is closely related to FOMAML and is simpler to implement. Unlike FOMAML, Reptile does not require a training-test split for each task, making it more natural in certain settings. The paper evaluates Reptile on the Mini-ImageNet and Omniglot datasets, showing that it performs slightly better than FOMAML on Mini-ImageNet and slightly worse on Omniglot. It also demonstrates that Reptile converges to a solution that is different from the minimizer of the expected loss. Theoretical analysis shows that Reptile's performance is driven by maximizing the inner product between gradients from different minibatches on the same task, which improves generalization. The paper also explores the effectiveness of Reptile in different settings, including the transductive setting where the model classifies the entire test set at once. It shows that Reptile performs well in this setting, suggesting that further research should focus on its use of batch normalization during testing. The authors also compare different inner-loop gradient combinations and find that Reptile benefits from taking many inner-loop steps, which is consistent with the optimal hyper-parameters found in previous experiments. Overall, the paper provides a comprehensive analysis of first-order meta-learning algorithms, showing that Reptile is a promising approach that can be implemented simply and effectively. The results suggest that Reptile is a strong baseline for meta-learning in various machine learning problems.