GANS FOR INCREMENTAL LEARNING

Classical machine learning models require all the data to be available prior to the training. Once a model is trained on certain tasks, it is limited to perform on those tasks and cannot be adopted to solve a different problem. But in the dynamic world that we are living, there is no end to a stream of data coming from different sources. Incremental learning or lifelong learning is a hot research area that studies the ability of neural networks to continuously learn new knowledge while retaining old experiences, imitating the human way of learning. One of the proposed approaches is to rehearse previously learned information when learning a new class of samples. In this work, we extract features embeddings from trained samples to use them for rehearsal instead of images themselves, or generated images from those classes. The benefit of this strategy is feature embeddings have less memory footprint, which is an essential attribute for memory critical applications. Furthermore, the feature vector of an image preserves considerably less sensitive data in itself than a raw image, subsiding privacy concerns. For this representation learning task, we propose a combination of three neural network paradigms, namely BiGAN, StyleGAN, and CGAN. BiGAN is a bidirectional GAN model in which the discriminator takes input from both the generator and encoder network. While the generator model serves its original task of generating realistic images, the encoder learns the inverse mapping from real data distribution to the feature representation. In the original implementation of BiGAN authors used DCGAN architecture which fails to generate higher-resolution images. In this project, more advanced network architectures, StyleGAN and ResNet have been adopted for generator and encoder sub-networks, respectively. Conditional GAN addresses the stochastic nature of generative adversarial networks and gives the possibility of generating samples from the desired class. At the end of the training process, the encoder sub-network is detached from the model and used for representation learning. Learned low-dimensional representations are later used for a classification task in the incremental learning settings and produced comparable results up against approaches based on raw images. Specifically, we compare the average incremental accuracy of our model against iCaRL, one of the most successful architectures developed for lifelong learning scenarios.