Autoencoder Aided Semi-supervised Incremental Learning of Handwritten Characters

Incremental learning is an important feature of next generation machine learning systems - it is about learning new classes incrementally by training on labeled images of the new classes. When the availability of the labeled images is limited, unlabeled images that are easily available, can be utilized in image classification using semi-supervised learning techniques. We demonstrate for the first time, the semi-supervised learning together with the incremental learning in neural network aided by auto-encoder. At first, an under-complete convolutional auto-encoder is trained on the unlabeled images of all the anticipated classes. Then, the pretrained encoder acting as feature extractor, is integrated with the classifier layers constructing a custom VGG network. Subsequently, the VGG model is trained on the labeled images incrementally in task incremental fashion keeping the feature extractor parameters unchanged. In our experiments on the EMNIST (extended MNIST) dataset, we obtain incremental learning accuracy of more than 96 % when the model is trained on 2000 labeled images for each class in the incremental learning phase. Due to the pretraining with the unlabeled images, our model performs well, even with fewer labeled images for training in the incremental learning phase; we obtain accuracy of more than 92 % even when the number of labeled images is reduced to 500 images for each class. We also present experimental results for the case, where the autoencoder is pretrained on randomly cropped images instead of the full images, thus learning from the partial features of the images. Despite using images that are cropped, we obtain incremental learning accuracy comparable to the result obtained for full images. The method proposed here may be useful in next generation machine learning applications due to the dual benefits of incremental learning and semi-supervised learning.