On First-Order Meta-Learning Algorithms

8 Mar 2018Alex NicholJoshua AchiamJohn Schulman

This paper considers meta-learning problems, where there is a distribution of tasks, and we would like to obtain an agent that performs well (i.e., learns quickly) when presented with a previously unseen task sampled from this distribution. We analyze a family of algorithms for learning a parameter initialization that can be fine-tuned quickly on a new task, using only first-order derivatives for the meta-learning updates... (read more)

PDF Abstract

Evaluation Results from the Paper


TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK COMPARE
Few-Shot Image Classification Mini-ImageNet - 1-Shot Learning Reptile + Transduction Accuracy 49.97% # 22
Few-Shot Image Classification Mini-ImageNet - 5-Shot Learning Reptile + Transduction Accuracy 65.99% # 19
Few-Shot Image Classification OMNIGLOT - 1-Shot Learning Reptile + Transduction Accuracy 97.68% # 7
Few-Shot Image Classification OMNIGLOT - 5-Shot Learning Reptile + Transduction Accuracy 99.48% # 7