Print 17 comment(s) - last by YashBudini.. on Sep 27 at 10:19 PM

The system is based off human vision systems and will be applied to fully-autonomous vehicles

Researchers at Yale University and New York University have developed a new supercomputer that is capable of navigating a car more quickly and efficiently through the use of a human-based visual system. 

The supercomputer is called NeuFlow, and it was created by Eugenio Culurciello of Yale's School of Engineering & Applied Science along with Yann LeCun from New York University. Culurciello developed the human-inspired system while LeCun supplied the complex vision algorithms, which runs the neural networks for synthetic vision applications. NeuFlow's actions are based on the human visual system, acting as quickly and efficiently as a human when obeying traffic laws, distinguishing different objects from one another such as trees and buildings, and reacting to other drivers on the road. 

Culurciello and LeCun are looking to use this supercomputer as a way for cars to drive themselves. To do this, NeuFlow runs more than 100 billion operations per second only using a few watts of power, which is less than what's required to power a cell phone. NeuFlow exists on a single chip, making it no larger than a wallet, but it is more efficient and powerful than full-scale computers. Also, this system "processes tens of megapixel images in real time." 

"One of our first prototypes of this system is already capable of outperforming graphic processors on vision tasks," said Culurciello.

The development of fully-autonomous vehicle's will be a significant advancement in the world of human convenience and safety, and that's why NeuFlow isn't the only computer-driven system out there right now. 

In 2008, DailyTech went for a spin in the Chevrolet Tahoe DARPA Challenge vehicle, which is a fully-autonomous vehicle that won the DARPA 2007 Urban Challenge and is equipped with GPS, radar, video, laser and LIDAR sensors and inputs to recognize objects on the road. Its key sensor, velodyne, has 64 sensors in a wide array and is able to collect one million bits of data per second at 10 Hz. It's logic consists of over 350,000 lines of code, and is able to obtain a 3-D view of the surrounding terrain just like NeuFlow.

But unlike the Chevrolet Tahoe DARPA Challenge vehicle, NeuFlow is not quite ready for vehicle use yet. Culurciello and LeCun are looking to use NeuFlow in other applications as well, such as a tool for 360-degree synthetic vision for soldiers in combat and to help improve robot navigation in dangerous locations. 

NeuFlow was presented by Culurciello at the High Performance Embedded Computing (HPEC) workshop in Boston, Mass. on September 15. 

Comments     Threshold

This article is over a month old, voting and posting comments is disabled

By KIAman on 9/16/2010 2:21:48 PM , Rating: 3
This is an amazing achievement in technology but the program can only be as good as their maker; humans. Humans are *gasp* prone to making errors.

Assuming the code is bug free, imagine the decision making logic a device like this needs. What thresholds need to be met to swerve from an object vs crashing into it? Do you swerve left or right? What happens when going over a sharp crest where you cannot see anything over it until you are over? etc.

I predict this company will start getting into AI or a fuzzy logic system.

RE: Logic
By Jaybus on 9/16/2010 2:51:11 PM , Rating: 2
It will have many advantages over a human driver. Obviously, in addition to visible light vision, it can have IR night vision, as well as distance and speed info via LIDAR/RADAR, and all at the same time. Unlike the human, it won't get itself into these situations as often in the first place, since it won't get drunk, become distracted, or make a choice to go over a sharp crest at extreme speeds. There is no question that nearly all accidents are caused by driver error and carelessness.

Anyway, the principle has already been shown in aircraft, which are clearly safer when on autopilot.

RE: Logic
By xthetenth on 9/16/2010 2:56:55 PM , Rating: 3
What thresholds need to be met to swerve from an object vs crashing into it?

Whether swerving can be done without something overly stupid happening such as hitting a tree. Just look to see if there's somewhere safe to swerve, if so, do it. If not, try to determine whether the objects off road or on road are safer to hit. That's the real tricky bit. Still, anticipation is the real key, and something an always alert computer could do much better than a human.

Do you swerve left or right?

Look at the object, calculate it's velocity and acceleration, figure out where it'll be when you pass, and go to the side which gives bigger clearance. Check the sides to make sure you aren't boxed in and you're good.

What happens when going over a sharp crest where you cannot see anything over it until you are over?

Maybe feather the accelerator, but other than that, it's not like humans do much more, and computer reflexes are already an advantage.

The real tricky bit is going to be building an object library so that it can make good guesses about objects' properties, such as weight and whether they're attached to the ground.

RE: Logic
By magneticfield on 9/17/2010 3:38:46 AM , Rating: 2
I think there are a lot of computer programs made by humans out there, running on computers made by humans, and which do things way better than humans.

This kind of software solution is called an "expert system".

Or take the example of the Japanese Automatic Train Control. You know their trains have a precise schedule, with a cumulated delay of a few seconds per year. Do you think humans could do better?

"Folks that want porn can buy an Android phone." -- Steve Jobs

Copyright 2016 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki