ROOT provides via the TMVA package a new Deep Learning framework, which supports convolutional and recurrent layers in addition to dense layers. The models can be trained also on NVIDIA GPU using CUDA. The CUDA implementation for convolutional layers is not fully optimized for Nvidia graphics cards. A possible way to improve its performance is to use the NVIDIA dedicated library for training and evaluating deep learning models, cuDNN.
The main goal is then to re-implement the GPU backend for convolutional network using this library from NVIDIA. As a further extension, capable candidates will also gain experiences from implementing a cuDNN rucurrent layers.
C++ skills, experience with large scale software development and GPU programming is a plus.