Meta-Learning for One-Shot Classification

Abstract

In this thesis, an ensemble method to improve one-shot classification was proposed. Ensemble method in deep learning usually is done by combining dif ferent initialization of prediction models with additional parameters as much as the number of ensemble members, whereas our method ensemble from the end points of the model in learning procedure without any additional parameters. This method can be applied in one-shot learning training phase which dataset is continuously being changed from episode to episode and end point has much more variance than usual classification problem. From this way, some ensemble members are better representing and classifying one-shot data, we find novel way to ensemble which is inspired by meta recognition system; meta-classifier. Meta classifier selects the best ensemble predictor and we achieve better performance than existing method on miniImageNet and CIFAR dataset. Furthermore, we ex perimented one-shot task with state-of-the-art base classifier and verify the best performance of one-shot classification.