Deep Learning Workshop

Deep Learning is rapidly penetrating into many scientific disciplines, as data generation and collection become easier while parallel computation power (especially in the form of GPU) grows stronger. LRZ deployed several GPU systems for deep learning, including a DGX-1 and Openstack cloud based GPU virtual servers (with Pascal 100) since the start of 2017. Together with Nvidia, we are organizing a one-day deep learning workshop on Sept. 14^th, 2017. This workshop is targeted at scientists with limited knowledge of Deep learning, but it will also be informative for Deep learning experts. The core of the workshop will be lectures given by Nvidia experts.  A preliminary program is shown as the following:

0900-0915 Welcome - Reinhold Bader (LRZ, HPC group leader)

0915-0930 NVIDIA Intro - Carlo Ruiz (NVIDIA, DGX EMEA Sales Director)

0930-1030 NVIDIA GPU Computing and DGX AI Supercomputer - Carlo Nardone (NVIDIA, Sr Solution Architect EMEA)

1030-1045 Changing gear for accelerating deep learning - the first year operation experience with DGX-1 - Yu Wang (LRZ)

1045-1100 break

1100-1200 Deep Learning Intro: AI boosting HPC - Carlo Nardone

1200-1230 DGX-1 simple demo - Carlo Nardone and Jonas Loof (NVIDIA, Deep Learning Solution Architect EMEA)

1230-1330 Lunch break

1330-1400 Invited talk: Bottlenecks towards Scalable Deep Learning on HPC Systems (Martin Kuehn, Competence Center High Performance Computing, Fraunhofer ITWM)

Abstract: The ability to scale the training of deep neural networks (DNN) to hundreds or even thousands of GPUs would result in great benefits for research and application of deep learning. However, distributed parallel training of DNNs is a very difficult and largely unsolved problem. In this talk, we outline the main theoretical and practical bottlenecks of DNN training on HCP systems and show results of our recent research, including distributed novel optimization methods, single sided communication, parallel data I/O and communication quantization.

1400-1430 Invited talk: Reconstructing and Understanding Indoor Environments using 3D DeepLearning  (Matthias Nießner, Computer Vision, TUM) 

Abstract: In the recent years, commodity 3D sensors have become easily and widely available. These advances in sensing technology have spawned significant interest in using captured 3D data for mapping and semantic understanding of 3D environments. In this talk, I will give an overview of our latest research in the context of 3D reconstruction of indoor environments. I will further talk about the use of 3D data in the context of modern machine learning techniques. Specifically, I will highlight the importance of training data, and how can we efficiently obtain labeled and self-supervised ground truth training datasets from captured 3D content. Finally, I will show a selection of state-of-the-art deep learning approaches, including discriminative semantic labeling of 3D scenes and generative reconstruction techniques. 

1430-1530 Multi-GPU applications: Deep Learning on DGX-1 - Jonas Loof

1530-1545 break

1545-1615 DGX-1 advanced demo - Jonas Loof

1615-1715 LRZ talks from workshop participants

1715-1730 Final remarks and conclusions

Note: The workshop participants are prefered to submit a three pages presentation about their research in deep learing, but this is not obligatory. Page 1, title, applicant names, background figure, 3 keywords (general discipline, speicial research field, deep learing software); Page 2, the results obtained so far on LRZ GPU resources or research they plan to carry out on LRZ GPU resources; Page 3, a wish list on LRZ GPU resources (more GPU hardware, longer reservation time, etc).  Please send your presentation (in powerpoint format) to wang@lrz.de. Based on the presentations, some participants will be selected to present their research at the workshop and be awarded extra computational time on DGX-1. The powerpoint presentation file name should be in the following format: firstname_lastname_dlw17.ppt.

Here are the presentations from Nvidia experts: Nvidia presentation.