Autoencoders neural networks optimization by visual analytics data

Abstract

Artificial Neural Networks, specifically those with a deep architecture (with more than two hidden layers), are usually affected by an occurrence known as super-training of data. In the super-training of data, the premature convergence of the network generates quite satisfactory results in the training and validation sets, but with a lower performance expected on the test set. Among the many ways of handle with this problem, such as regularization and use of larger training sets, one has been rarely employed is the one that makes use of information visualization techniques. These techniques incorporate the user's knowledge into the network training. Such visualization tools enable a greater understanding of the network learning process, allowing to intervene in a certain region, parameters, connections and neurons of the network in order to improve its effectiveness. Depending on the network configuration, the user can check how the samples are grouped in the final classification process, as well as visualize the weights and activations of the neurons and how they affect the final classification result. In this sense, the present proposal of a master's research project aims to use information visualization techniques to assist in the training process of Autoencoder Neural Networks, since this technique has been widely used in the literature for the most diverse tasks, such as classification of images, signals and dimensionality reduction. The present proposal also contemplates a period of internship abroad by BEPE (Bolsa de Pesquisa e Estágio no Exterior) at the University of Groningen, Netherlands, under the guidance of Prof. Alexandru Telea, who has experience in the information visualization area. Given the student already has previous experience in deep learning due to her scientific initiation supported by FAPESP, this internship abroad will contribute to her learning in the data visualization area, as well as allowing a scientific maturity and experience with other research groups. (AU)