On First-Order Meta-Learning Algorithms

8 Mar 2018  ·  Alex Nichol, Joshua Achiam, John 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. This family includes and generalizes first-order MAML, an approximation to MAML obtained by ignoring second-order derivatives. It also includes Reptile, a new algorithm that we introduce here, which works by repeatedly sampling a task, training on it, and moving the initialization towards the trained weights on that task. We expand on the results from Finn et al. showing that first-order meta-learning algorithms perform well on some well-established benchmarks for few-shot classification, and we provide theoretical analysis aimed at understanding why these algorithms work.

PDF Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Few-Shot Image Classification Mini-Imagenet 10-way (1-shot) Reptile Accuracy 31.1 # 14
Few-Shot Image Classification Mini-Imagenet 10-way (1-shot) Reptile+BN Accuracy 32.0 # 11
Few-Shot Image Classification Mini-Imagenet 10-way (5-shot) Reptile Accuracy 44.7 # 14
Few-Shot Image Classification Mini-Imagenet 10-way (5-shot) Reptile+BN Accuracy 47.6 # 12
Few-Shot Image Classification Mini-Imagenet 5-way (1-shot) Reptile + Transduction Accuracy 49.97 # 78
Few-Shot Image Classification Mini-Imagenet 5-way (5-shot) Reptile + Transduction Accuracy 65.99 # 67
Few-Shot Image Classification OMNIGLOT - 1-Shot, 20-way Reptile + Transduction Accuracy 89.43% # 17
Few-Shot Image Classification OMNIGLOT - 1-Shot, 5-way Reptile + Transduction Accuracy 97.68 # 17
Few-Shot Image Classification OMNIGLOT - 5-Shot, 20-way Reptile + Transduction Accuracy 97.12% # 17
Few-Shot Image Classification OMNIGLOT - 5-Shot, 5-way Reptile + Transduction Accuracy 99.48 # 13
Few-Shot Image Classification Tiered ImageNet 10-way (1-shot) Reptile Accuracy 33.7 # 13
Few-Shot Image Classification Tiered ImageNet 10-way (1-shot) Reptile+BN Accuracy 35.3 # 10
Few-Shot Image Classification Tiered ImageNet 10-way (5-shot) Reptile Accuracy 48.0 # 13
Few-Shot Image Classification Tiered ImageNet 10-way (5-shot) Reptile+BN Accuracy 52.0 # 12

Results from Other Papers


Task Dataset Model Metric Name Metric Value Rank Source Paper Compare
Image Classification Tiered ImageNet 5-way (5-shot) Reptile Accuracy 66.47 # 7
Image Classification Tiered ImageNet 5-way (5-shot) Reptile + BN Accuracy 71.03 # 3

Methods