Deep Learning Architectures for Hard Character Classification

Recent research indicates that deep learning has achieved noticeably promising results in a wide range of areas such as computer vision, speech recognition and natural language processing. This paper offers an empirical study on the use of deep learning techniques for hard characters recognition on the notMNIST dataset. The MNIST dataset has been widely used for training and testing in the field of machine learning such as for the performance comparison of different deep learning algorithms. However, similar performance evaluation using the notMNIST dataset has not been reported. This dataset is harder and much less clean than the MNIST dataset. In this paper, we constructed several experiments to evaluate various deep learning architectures and proposed a multi-layer convolutional neural network for large-scale hard character classification on the notMNIST dataset. The result shows that our method can achieve 98% accuracy of classification. Comparisons were also performed against conventional fine tuning models such as logistic classifier and shallow neural network to demonstrate that well-constructed deep neural networks can significantly improve the accuracy of hard character classification on the notMNIST dataset.