NYU Center for Data Science Hosts NVIDIA Day
NVIDIA researchers discuss the future of AI computing and autonomous driving
Last week, we welcomed NVIDIA to spend a whole day with our students. A world leader in visual computing technologies, NVIDIA is a $7 billion hardware and software company with 10,000 employees.
Since the company’s founding in 1993, they’ve been leaders in the gaming market. They invented the graphics processing unit in 1999, and since then, have brought the parallel processing power of GPU computing to areas such as virtual reality, artificial intelligence, and self-driving cars.
Joan Bruna, CDS faculty member and Assistant Professor of Computer Science and Data Science, kicked off NVIDIA Day with a presentation of his research on “divide and conquer networks.” These are neural networks that control computational complexity by splitting and merging inputs, allowing them to solve problems that are computationally or statistically hard. Applications for these networks include optimizing transportation costs and resource allocation.
Nikolai Yakovenko, a Sr. Research Scientist in the Applied Deep Learning Research group at NVIDIA, followed with an overview of the company’s AI research and development. He focused on five examples of applied deep learning research topics:
- Assisted game design: AI systems can encode emotion from a voice actor’s audio performance into the digital rendering of a character’s face.
- Next-generation game rendering: AI can use ray-tracing, a way of rendering computerized images by tracing the path of natural light, to produce computerized images that are almost indistinguishable from photographs or video footage.
- Synthetic data: AI systems can be trained with synthetic data to identify specific objects like cars.
- Audio de-noising: White noise or other disruptive background sounds can be removed from audio recordings by AI systems.
- Candidate resume screening: AI can rank candidates in tiers based on certain criteria, and it can predict how far candidates will progress towards a job offer.
Mariusz Bojarski, a deep-learning R&D engineer at NVIDIA, concluded NVIDIA Day by explaining their autonomous driving system. At their lab in Holmdel, NJ, Bojarski’s team has taught a unique convolutional neural network (CNN) to drive a car by observing human drivers and emulating their behavior. The CNN, called PilotNet, relies on a center-mounted camera on the vehicle and two mirror-mounted cameras on each side.
NVIDIA evaluates its autonomous driving system based on three performance metrics: autonomy (the percentage of the time the car drives without human intervention), precision (the percentage of the time the car deviates from what a human driver would do), and comfort (measured based on the smoothness of steering).
Future challenges for NVIDIA’s autonomous driving team include transferring neural networks among different cars, enhancing the sensor capabilities with additional cameras, and enabling PilotNet to predict the behavior of other drivers on the road. Bojarski hopes to see PilotNet reach the benchmark of driving one million miles without human intervention.
The big day ended with NVIDIA raffling off an NVIDIA TITAN Xp — a powerful GPU for gamers and creatives — to one of our lucky students.
By Paul Oliver