Learning to Drive One Selfie at a Time
I’ve just returned from a visit to sunny California where I saw lots of exciting innovations Silicon Valley is preparing for the world. But nothing impressed me more than what I saw at Tesla. While at company headquarters, I got to see firsthand how they’re poised to disrupt the car market and revolutionize driving for everyone — while also becoming the most dominant carmaker in the world. How? By finding a new approach to designing self-driving cars. To appreciate how visionary Tesla’s new method is, however, I need to tell you a little bit about the wrong methods they and others have been trying.
For years car makers have been designing self-driving cars through a traditional coding strategy called decomposition. In decomposition, a designer identifies a single task — in this case, driving — and breaks it down into a series of smaller tasks — parking, steering, avoiding collisions, etc. Decomposition has been successfully deployed in all the life-changing technology that has emerged for decades, so there was every reason to think this method was adequate for self-driving cars. But then Tesla discovered there were certain problems decomposition couldn’t handle.
“Building a machine, that is beating a human driver with 99.9999999% (nine 9s) is only possible with an unfair advantage. The answer is networked autonomy. Real-time data network effects help machines know without seeing.”
Consider all that an automated car must be able to do. It must navigate construction sites, every one of which is slightly different. It must recognize stop signs, some of which are partially obscured or poorly illuminated. It must be able to tell the difference between a dog, a baby, and a doll. It must also be able to do these and other tasks at night. Though humans do have accidents, they’re remarkably skilled at navigating these challenges, getting them right 99.9999999% of the time. For computers, these tasks pose real trouble.
The problem proves to be one of vision. Think about all the different ways a person can look — old or young, tall or short, sitting or standing, angry or happy. A brain has no trouble seeing all these variations as a “person,” but using decomposition to get a computer to do so would require breaking down the category of “person” into countless smaller subsets of people. That not only takes forever, it doesn’t work very well. Then Tesla had a breakthrough. They realized that the trick was not to focus so much on the countless small tasks but to focus on the optimization of the few tasks they assign the machine. Let’s understand what that means in practice.
In the example of the person, engineers upload countless photos of different people in different poses and identify each photo as a “person.” When this process is finished, the software has a huge database of highly variable images, all of which the software knows are people. Then, when the car “sees” an object and wants to identify it, the computer will scan its inventory of images to decide if what it sees closely matches any of the “people” it has on file. If what the car sees does match an image, the car will respond to the person.
Of course, driving is a lot more complicated than identifying people, but Tesla has discovered that optimizing the software through vast collections of images is an effective approach to all aspects of self-driving cars. In fact, the optimization process is proving so successful that the company will soon be testing its cars on a coast-to-coast drive across America. Amazingly, the software networks that Tesla has designed over the past years will soon be more powerful than the brain that evolved over millions of years. In short, the software will soon see things that people don’t see, which means these networks will transform a wide range of industries.
Tesla’s success must serve as a wake-up call for other companies, all of which are locked in a traditional approach of decomposition and are therefore nowhere close to being road-ready. For make no mistake: the future of driving is automation. The self-driving car is one of the few inventions that will make life better for everyone: cars will be more reliable for passengers, greener for the planet, and safer for the more than a million people who annually die on the world’s roadways.
Even car manufacturers with strong brands and loyal followers will find it hard to compete with cars that save the lives of drivers and their families. That’s why the companies that lose the race to develop these cars will suffer worse than Detroit did in the 1980s. Further, Tesla’s optimization method shows that the time for companies to invest is now. Because optimization requires so much time and patience to develop, it will be impossible to play catch-up once these self-driving cars hit the market.
Yet, I am excited. Networked autonomous cars are going to make the world a better place, and the optimization techniques will be applicable to so many fields that more exciting innovations are sure to come. And even more exciting is that there are still so many areas of research left to perfect that any company could still conquer this market and become its leading brand. It’s already safe to say, however, that the company who drives across the finish line first tomorrow will have done so through a patient process of optimization undertaken today.
