Deep learning for temperature prediction using GPUs

Abstract

Deep learning techniques have proven to be very successful when dealing with highly-complex problems such as speech and image recognition. Generally, deep architectures are composed of several layers of nonlinear transformations that are intended to capture highly-complex relationships. This project was intended to analyse how deep architectures could be used for tackling temperature prediction. Furthermore, this thesis was intended to analyse how Graphical Processing Units(GPUs) could increase performance over CPUs. For this undertaking a python library named Theano was used. This library allowed an easy GPU utilisation. One of the problems that arise when training deep architectures is that they are difficult to train because of several problems related to these architectures, such as gradient diffusion and non-convexity. A commonly used solution is to pre-train the Neural Networks to reduce these problems. After training several different models, it was not possible to develop a deep architecture that could outperform shallow ones. GPUs generally performed better than CPUs. It was shown how the larger the Neural Networks, the better the GPU over CPU performance was.