TripletMAML: A metric-based model-agnostic meta-learning algorithm for few-shot classification

Abstract

In this paper, we introduce TripletMAML, a new meta-learning algorithm that enhances the Model-Agnostic Meta-Learning (MAML) approach by incorporating a metric-learning dimension. This enhancement involves the adoption of MAML's optimization strategies while transitioning to a triplet network model to facilitate metric learning. A novel aspect of this approach is our triplet-task generation technique, designed to produce meta-learning tasks with triplets for both 1-shot and 5-shot settings. TripletMAML extends MAML by jointly incorporating metric-learning and optimization-based principles through a triplet-task formulation, offering a unified and effective framework for few-shot classification. We evaluate TripletMAML's effectiveness across four well-known few-shot image classification benchmarks, comparing its performance against a range of baseline methods. Our findings indicate that TripletMAML, even without data augmentation or extensive hyper-parameter adjustments, significantly improves MAML's performance and surpasses competing baseline approaches in both 1-shot and 5-shot settings. We also demonstrate that optimizing the hyper-parameters automatically using differential evolution method can elevate TripletMAML's performance to that of more sophisticated models. Additionally, we conduct image retrieval experiments to ascertain whether TripletMAML's few-shot classification training provides a good starting point for addressing few-shot image retrieval challenges. The source code for our study is available at https://github.com/aylagulcu/TripletMAML.