Reptile: A scalable meta-learning algorithm

Researchers have developed Reptile, a new meta-learning algorithm that improves machine learning efficiency by repeatedly sampling tasks and updating parameters through stochastic gradient descent. The algorithm is mathematically similar to first-order MAML but requires only black-box access to optimizers like SGD or Adam while maintaining similar performance and computational efficiency.

• →Reptile is a simple meta-learning algorithm that uses repeated task sampling and parameter updates.
• →The algorithm applies the Shortest Descent method to meta-learning scenarios.
• →It only requires black-box access to standard optimizers like SGD or Adam.
• →Performance and computational efficiency are similar to first-order MAML.
• →The approach simplifies the meta-learning process while maintaining effectiveness.