Page Comparison

etc.cuit.columbia.edu

Urs Muller

Dir. of Developer Technology & Chief Software Architect, NVIDIA

The Future of Autonomous Vehicles

Urs Muller's work focuses on the development of end-to-end solutions for autonomous vehicles. He has 20+ years of experience in robotics, computer vision, machine learning, and high-performance computing. He received Ph.Ds in Electrical Engineering and Computer Science from the Swiss Federal Institute of Technology in 1993.

NVIDIA - founded 1993, 11,000 employees, based in Santa Clara, CA

focus on gaming, VR/AR, Data Center, Self driving cars

all initiatives linked by the same underlying AI and visual computing architecture

Drive-AGX stack:

• Drive AV
• Driveworks SDK
• Drive OS
• Drive AGX Xavier and Pegasus Hardware
  • • Xavier - 1 processor, 30 TOPS for deep learning, 30 watts power
    • Pegasus - 2 Xavier processors,

NVIDIA rents space on former Bell Labs campus in Holmdel, NJ

Bell Labs pioneered Deep Learning in 80's and 90's

Deep learning allows us to solve problems that we don't know how to program

led to development of neural networks

Bell Works campus is ideal to test self-driving cars due to extensive private road network

Convolutional Neural Networks access prior learning

NVIDIA view of learning:

• choose the right structure
• allow machines to access prior knowledge

Lessons learned:

• look at the data during all processing steps
• solid debugging tools critical
• validate the training data
• the work is experimental
• work with real data

1995 - deployed Holmdel neural nets at Wachovia bank; eventually processed 20% of checks in US

1995 - AT&T Bell labs fractured

2002 - AT&T mass layoffs

2003 - Bell Labs work led to new programs at DARPA

2012 - Deep Learning becomes more popular and commercially potentially viable in commercial applications

NJ Team Goal: solve the hard and unsolved autonomous vehicle problems using machine learning

simple driving rules fail due to numerous variable road conditions and unwritten behaviors

hard to write down every single rule of safe driving; relatively easy to collect training data through driving

training data is gathered through pixel input (sensors) during actual human-driven path. Convolutional Neural Network prediction of path is compared to human path, and discrepancy is provided as feedback to the CNN for additional learning

Open challenges:

deal with ambiguous situations - there is often more than one correct answer (4-way stop signs)

learn from imperfect behavior (several observations; most correct; some not correct)