Mathematical analysis on meta-learning capability of memory augmented neural networks

Abstract

Recently, Machine Learning technology using Deep Learning is achieving remarkable successes. Deep Learning shows good result in performance, but it needs a lot of data and computing resources. The One-Shot Learning algorithm is an way to break through this limit of deep learning. As an way to realize one-shot learning, meta learning using memoryaugmented neural network was proposed. In this thesis, we will find the origin of the meta learning capability of memory augmented neural network using various experimental and mathematical analysis. Also, we will find the way to improve the performance of memory augmented neural network based on these analysis.