Triplet MAML for Few-shot Classification Problems

Abstract

In this study, we propose a TripletMAML algorithm as an extension to Model-Agnostic Meta-Learning (MAML) which is the most widely-used optimization-based meta-learning algorithm. We approach MAML from a metric-learning perspective and train it using meta-learning tasks composed of triplets of images. The idea of meta-learning is preserved while generating the meta-learning tasks and training our novel meta-model. The experimental results obtained on four few-shot classification datasets show that TripletMAML that is trained using a combined loss yields in high quality results. We compared the performance of TripletMAML to several metric learning-based methods and a baseline method, in addition to MAML. For fair comparison, we used the reported results of those algorithms that were obtained using the same shallow backbone. The results show that TripletMAML improves MAML by a large margin, and yields better results than most of the compared algorithms in both 1-shot and 5-shot settings. Moreover, when we consider the classification performance of other meta-learning algorithms that use much deeper backbones, we conclude that TripletMAML is not only competitive in terms of the classification performance but also very efficient in terms of the complexity.