Deep Image Restoration, Vision and Reinforcement Learning
06-23-2017 at 11.00a.m.
Aula G09, Via Golgi 19 Milan
Speaker: Iuri Frosio, Senior Research Scientist, NVIDIA
Responsible: Alberto Borghese
In the last few years, the availability of large amount of data together with the computational power offered by NVIDIA GPUs and the re-discovery of neural networks have driven the machine learning revolution. The advent of deep learning has dramatically changed our approach to image processing, computer vision and reinforcement learning problems, producing new state-of-the-art results in most of these fields. Being part of NVIDIA research means having a privileged point of view on these events. In the first part of the talk I will briefly summarize some of the recent results achieved by the Visual Computing and Machine Learning Research group in NVIDIA, ranging from hand gesture recognition in automotive to the pruning of deep neural networks. I will also describe in detail how deep neural networks can be employed for general image restoration tasks and used as a valid alternative to the traditional Image-Signal-Processing based pipeline, widely implemented in nowadays cameras. In the second part of the talk, I will focus my attention on deep reinforcement learning, and in particular to the efficient implementation of the Advantage-Asynchronous-Actor-Critic (A3C) algorithm on a GPU, showing how to achieve a 45x speed up compared to a CPU implementation of the same algorithm.