Neural Architecture Search Using Differential Evolution in MAML Framework for Few-Shot Classification Problems

Abstract

Model-Agnostic Meta-Learning (MAML) algorithm is an optimization based meta-learning algorithm which aims to find a good initial state of the neural network that can then be adapted to any novel task using a few optimization steps. In this study, we take MAML with a simple four-block convolution architecture as our baseline, and try to improve its few-shot classification performance by using an architecture generated automatically through the neural architecture search process. We use differential evolution algorithm as the search strategy for searching over cells within a predefined search space. We have performed our experiments using two well-known few-shot classification datasets, mini-ImageNet and FC100 dataset. For each of those datasets, the performance of the original MAML is compared to the performance of our MAML-NAS model under both 1-shot 5-way and 5-shot 5-way settings. The results reveal that MAML-NAS results in better or at least comparable accuracy values for both of the datasets in all settings. More importantly, this performance is achieved by much simpler architectures, that is architectures requiring less floating-point operations.