Multi-task Learning on MNIST Image Datasets

We apply multi-task learning to image classification tasks on MNIST-like datasets. MNIST dataset has been referred to as the {\em drosophila} of machine learning and has been the testbed of many learning theories. The NotMNIST dataset and the FashionMNIST dataset have been created with the MNIST dataset as reference. In this work, we exploit these MNIST-like datasets for multi-task learning. The datasets are pooled together for learning the parameters of joint classification networks. Then the learned parameters are used as the initial parameters to retrain disjoint classification networks. The baseline recognition model are all-convolution neural networks. Without multi-task learning, the recognition accuracies for MNIST, NotMNIST and FashionMNIST are 99.56\%, 97.22\% and 94.32\% respectively. With multi-task learning to pre-train the networks, the recognition accuracies are respectively 99.70\%, 97.46\% and 95.25\%. The results re-affirm that multi-task learning framework, even with data with different genres, does lead to significant improvement.