With this thesis, we investigate the use of FPGAs as an alternative hardware platform and we propose an architecture for accelerating critical operations (convolutional layers) while maintaining full compatibility with Caffe deep learning framework.