Convolutional Deep Neural Networks on GPUs for Particle Physics Applications

The project's ultimate goal is to provide a GPU implementation for the existing Convolutional Neural Network package within root/tmva. During my preliminary work with the codebase, I discovered that the current package's public interface can be further improved. Since performing this change can significantly reduce the complexity of my main task, as well as any future extensions to the package, I plan to work on it during the first phase of the summer period. The first part of the present proposal goes through the necessary changes to achieve a clean API. The rest of the proposal iterates through the main modules that need to be ported into the GPU implementation. These are the different layer types that can be included in a convolutional neural network, as well as a number of generic helper functions. Attention is drawn on guaranteeing the deliverable's quality, both in terms of correctness and in terms of speed-up. This will be achieved through extensive testing and standardized benchmarking respectively.