# How to Train Your MAML to Excel in Few-Shot Classification

**Authors:** Han-Jia Ye, Wei-Lun Chao  
**arXiv:** [2106.16245](https://arxiv.org/abs/2106.16245)  
**Date:** 2021-06-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2106.16245v1/

## Key points

- Model-Agnostic Meta-Learning (MAML) algorithm faces challenges in few-shot classification tasks
- Need for a large number of gradient steps during inner loop update to enhance MAML's performance
- Sensitivity to permutation of class assignments during meta-testing, leading to variance
- Learning a shared classifier initialization for all classes improves results
- UNICORN-MAML approach achieves competitive performance on benchmark datasets like MiniImageNet and TieredImageNet
- UNICORN-MAML maintains simplicity of MAML while achieving comparable or superior results to state-of-the-art algorithms

## Summary

The paper "How to Train Your MAML to Excel in Few-Shot Classification" delves into the challenges faced by the Model-Agnostic Meta-Learning (MAML) algorithm when applied to few-shot classification tasks. MAML is a versatile meta-learning algorithm that can adapt to various model architectures and problem domains. However, its performance in few-shot classification lags behind more specialized algorithms. The authors identify key factors that can enhance MAML's performance in this context. One crucial finding is the need for a large number of gradient steps during the inner loop update, contrary to common practices in few-shot learning. Additionally, MAML proves sensitive to the permutation of class assignments during meta-testing, leading to significant variance due to the exponential possibilities of assigning learned initializations to classes. To address this challenge, the authors explore methods for achieving permutation invariance and discover that learning a shared classifier initialization for all classes yields optimal results. The proposed approach, named UNICORN-MAML, showcases competitive performance on benchmark datasets like MiniImageNet and TieredImageNet. By incorporating these insights without introducing additional sub-networks, UNICORN-MAML achieves comparable or superior results compared to state-of-the-art algorithms while maintaining the simplicity of MAML. In terms of experimental setup, the paper details the datasets used (MiniImageNet and TieredImageNet), neural network architecture choices, and implementation specifics. Tasks involve sampling N-way K-shot tasks during meta-training, meta-testing, and meta-validation phases from designated classes and images within these datasets. Overall,"How to Train Your MAML" provides valuable insights into optimizing MAML for few-shot classification tasks through careful consideration of inner loop gradient steps and permutation invariance strategies. The proposed UNICORN-MAML approach demonstrates promising results on challenging benchmark datasets while preserving the fundamental simplicity of MAML's design.