Evolutive Deep Models for Online Learning on Datastreams with no Storage.

In recent years Deep Learning based methods gained a growing recognition in many applications and became the state-of-the-art approach in various fields of Machine Learning, such as Object Recognition, Scene Understanding, Natural Language processing and others. Nevertheless, most of the applications of Deep Learning use static datasets which do not change over time. This scenario does not respond well to a big number of recent applications (such as tendency analysis on social networks, video surveillance, sensor monitoring, etc.), especially when working with data streams which require real-time adaptation to the content of the data. In this paper, we propose a model that is able to perform online data classification and can adapt to data classes, never seen by the model before, while preserving previously learned information. Our approach does not need to store and reuse previous observations, which is a big advantage for data-streams applications, since the dataset one wants to work with can potentially be of very large size. To make up for the absence of previous data, the proposed model uses a recently developed Generative Adversarial Network to drive a Deep Convolutional Network for the main classification task. More specifically, we propagate generative models instead of the data itself, to be able to regenerate the historical training data that we didn't keep. We test our proposition on the well known MNIST benchmark database, where our method achieves results close to the state of the art convolutional networks trained by using the full dataset. We also study the impact of dataset re-generation with GANs on the learning process.