Estudo comparativo entre plataformas de deep learning

Abstract

Deep learning techniques has been showing advances in various Machine learning tasks. However, the implementation of these techniques is very complex. Thus, to help the implementation of Deep learning projects, software tools are being proposed. A considerable amount of these tools already exists. This leads to a difficulty on the choice of who is looking to start a project. In order, to assist in this choice, this work brings a comparative study between the open source and distributed Apache Singa, Graphlab and H2O platforms. Detailed test results using a database composed of images and another composed of alphanumeric attributes were produced. And aspects such as training time, prediction time, hardware resource utilization and accuracy of the algorithms of each platform were evaluated.