I’m finishing up at the NVIDIA GPU Developer Conference [Disclosure: NVIDIA is a client of the author]. The conference started out to be mostly about gaming, but now gaming is just a small part of the overall event, which has a massive professional VR, deep learning and automotive focus. The final keynote of the event was by Gill Pratt, CEO of Toyota Research. Toyota Research is not only heavily involved with the future of cars, but has been prolific with regard to robotic research. With the coming wave of autonomous vehicles coming to market this talk couldn’t have been more timely.
[ Related: Toyota funds AI research to build autonomous cars ]
These are what I pulled away as highlights.
This is the number of people who die every year from car accidents. It continues to shock me just how big this number is. The focus of the autonomous car effort is largely to eliminate these deaths.
Our brains are amazingly power-efficient
Current electrical autonomous car systems take thousands of watts of power to operate while our brains only use 30 watts of power and they can drive cars part time (we tend to daydream while driving). The technological challenge is to create a solution that can perform this same task within the same power envelope.
Apparently, right now, a robot that looks like a person, and a robot that looks like a horse are both a hundred times less power efficient than the real thing. What a massive amount of research has discovered is that nature is naturally very power efficient. To animals, energy is very expensive so evolution automatically optimizes for power. This suggests that modeling after nature is the most successful path to solving this problem.
Neovision, a vision product created before deep learning, was a visual system created to emulate nature and instead of the typical model where complexity is expensive and power was cheap, it worked on the model that complexity was cheap and energy was expensive and the result was thousands of times more efficient. Basically, it was highly specialized and parts not used were turned off.
DARPA Synapse, which was an exercise comparing human brains to Von Neumann Computers, further support and found that there was a massive disparity between complexity and power between the two systems. The conclusion was, for uses like autonomous driving, more hardware is better, but only if you can aggressively turn off what is not in use. The result will be smarter and more power efficient thinking machines.
[ Related: Ford will triple its autonomous car test fleet ]
One of the big problems with autonomous cars is the idea that when there is a problem the car will turn over control to an unprepared driver who likely will immediately crash. Apparently there are three types of autonomy. (I have this mental image of a driver who has been reading a book suddenly being handed control of the crashing car just getting out “oh cra…” before boom). The three types of autonomy are series, interleaved and parallel:
Series is an order-based system where the person gives the robot an order and it follows it.
Interleaved is shared control like between a pilot and co-pilot. In a car this would be a case where the car hands off control to the user when it is in trouble.
Parallel this is where both the human and the robot are working together all the time and the robot learns from the human over time until it can operate autonomously. Basically, this would be teaching a car to drive like you would teach a human. The example was of a quadriplegic operator connected to a robotic arm both with and without computer assist. Without computer assist it was largely unsuccessful while with computer assist it was successful every time and the user never noticed the difference and attributed the success to skill.
This has resulted in two models. The “chauffer” model where you get in the car and it just drives you operating 100 percent of the time and the “guardian angel” model. This second model is in use today and provides different levels of intervention, today you see it in anti-lock brakes and accident avoidance systems. The first system can’t make a mistake, the second just can’t make anything worse (and generally makes them much better).