One-Shot Digit Classification Based on Human Concept Learning

Abstract

One of the key challenges of present machine learning approaches is to match the human-level performance in terms of number of examples for training. Even though deep learning has achieved remarkable accuracy and speed for classification problems, the performance still depends on the number of examples used for training. In this paper, we explore the problem of classification of handwritten digits from a single training example using a probabilistic approach based on the process in which characters are generated and learned by humans. The results obtained suggest that understanding of the human process can help in achieving good classification results even with sparse data for training.