Incremental Learning of Handwritten Characters in EMNIST dataset

Abstract

An important characteristics of the brain is the ability to learn something new without forgetting the knowledge learned earlier. On the contrary, catastrophic forgetting of the previous tasks is observed in typical artificial neural networks when we try to fine-tune a trained neural network to train on additional tasks. Recently, various methods have been developed to alleviate the forgetting in neural networks enabling incremental learning of new tasks. In this work we investigate for the first time, the task incremental learning of handwritten characters in EMNIST (Extended MNIST) dataset. In our experiments, we have obtained the incremental learning accuracy of more than 96% in task incremental scenario. We present experimental results on the high accuracy incremental learning on EMNIST dataset along with the confusion matrices obtained using custom VGG networks. This may be useful in the development of lightweight machine learning apps supporting handwritten character recognition for languages with relatively large character set or multiple character sets required in a multilingual environment.