New Bug-Hunting Method for Self-Driving Cars

Home > News > Content

【Summary】DeepXplore is a new debugging method for deep learning algorithms. It has the potential to save lives by making autonomous cars safer.

Mia Bevacqua    Nov 18, 2017 4:15 PM PT
New Bug-Hunting Method for Self-Driving Cars

Microsoft Word crashes before a document is saved, or Skype cuts out during the middle of a conference call. Most software bugs are annoying, but not deadly. That is, unless you're dealing with self-driving cars. Flaws in autonomous vehicle software have the potential to cause accidents and fatalities. That's why researchers are working on a bug detection method, called DeepXplore, to find problems in deep learning algorithms. 

Nueral Network Debugging 

Deep learning is a form of artificial intelligence (AI) that allows autonomous vehicles to learn as they go. To do this, deep learning uses what's referred to as neural networks, which enable the vehicle to gather information from data images. Deep learning differs from traditional software that follows a fixed, "if this, do that" structure. Because it's so different, deep leaning has unique bugs that require an equally unique detection method.

DeepXplore uses at least three deep learning neural networks to combat bugs. It cross refences the networks with one another for accuracy. If it finds one of the networks has gone astray, it retrains that network so it can fall back in line. Researchers at Columbia University and Lehigh University worked together to create the method.

The DeepXplore Difference

Until DeepXplore, neural network debugging was rather tedious and inaccurate. According to IEE, one method involves researchers feeding test images into the networks until a wrong decision is made. Another, called adversarial testing, creates a series of test images until one is found that causes the network to falter. 

But DeepXplore takes a different approach. It creates test images that are most likely to make three or more neural networks make incorrect decisions. For example, it might look for a certain parameter that could lead two networks to identify a vehicle as a car, while the third identifies it as a human.

DeepXplore also activates many neurons and neural network pathways for the best coverage. This process was able to activate 100 percent of network neurons. That's about 30 percent more on average than either other test method used in deep learning. 

At this point, DeepXplore isn't guaranteed to find all the bugs in deep learning AI. It is, however, far more comprehensive than previous bug-hunting methods. It also helps get self-driving cars one step closer to public roadways, by making them safer. 

Prev                  Next
Writer's other posts
    Related Content