CDS Hosts Third Annual NVIDIA Day
NVIDIA returns to CDS with presentations about applied deep learning and autonomous driving
During the third annual NVIDIA Day at NYU’s Center for Data Science, NVIDIA researchers shared some of their recent projects with CDS students, including an update on their autonomous driving research. Kyunghyun Cho, Assistant Professor of Computer Science and Data Science and NVIDIA Partner, kicked off the day by highlighting the revolutionary role of NVIDIA’s GPUs within machine learning and neural network research. Cho emphasized CDS’ relationship with the NVIDIA AI Lab, and said, “We help NVIDIA realize the potential of their machines.”
NVIDIA’s Higher Education Research Account Manager Ronak Shah added that “NVIDIA creates new markets through specialization.” He introduced the NVIDIA presenters, which included Nikolai Yakovenko, Senior Deep Learning Research Scientist, Aysegul Dundar, Deep Learning Researcher for Autonomous Driving, and Adam Lesnikowski, Senior Software Perception Engineer.
Yakovenko discussed applied research at NVIDIA — work that is one to two years from completion, focused on concrete problems without solutions. His work often involves building prototypes, and sometimes continues until a product is fully developed. Yakovenko explained current projects involving synthetic video, text-to-speech, and unsupervised language modeling. NVIDIA’s frame prediction process uses deep learning supersampling, a marquee feature of its new Turing architecture GPUs, to improve image quality of predicted video frames. Yakovenko also works on text-to-speech systems and unsupervised language modeling problems. His team trained a model to label emotion with 40 GB of text, and their model outperformed comparable models from IBM Watson and Google in fine-tuned categories.
Dundar followed with an introduction of NVIDIA AI Infra, an industrial-grade deep learning platform for autonomous vehicles. This involves large-scale labeling and training — with petabytes of data. Dundar stated bluntly, “Building AI for self-driving cars is hard.” She’s especially interested in the domain adaptation problem, which means making sure autonomous vehicles are capable of driving in every physical and temporal environment, from a desert in the daytime to a tundra at night. Solving this problem requires real images from all kinds of locations and unsupervised image translation to account for windshield reflections. However, Dundar said models trained with large amounts of synthetic data saw improved results when tested with real data.
Lesnikowski concluded NVIDIA Day with a presentation about deep active learning in autonomous vehicle research. Two big challenges for his research are how to best spend time and money on labeling, and how to find the best examples in a dataset. For the autonomous vehicle problem, the most valuable data are outlier examples of disturbances on roads. This data comes from car sensors, and again amounts to petabytes. To achieve this labeling task, Lesnikowski’s team recruited a model to pass over unlabeled data and select viable data for labeling. This data then gets manually labeled and incorporated into the training set for the model, and the process continues cyclically. But to improve results, Lesnikowski started using two different models to select data for labeling from the same dataset.
Lesnikowski concluded the day by remarking that deep active learning is effective, their methods have led to significant efficiency gains, their models can find good examples, and there is a lot of exciting work happening at NVIDIA!
By Paul Oliver