Andrew has been writing Anti Buzz for 4 years resulting in almost 200 articles. For the next several weeks we will revisit some of these just in case you missed it. This article from early 2013 0n driverless cars is even more relevant today.
BIG NEWS, We may have to wait a bit longer for some new Anti Buzz articles as Andrew just became a new father. Congratulations. That means that I just became a Grandfather.
The Buzz: A computer can’t possibly account for all the variable in driving a car!
The Anti-buzz: Driving is only arithmetic.
Early in the year while we are still looking forward, I’d like to talk about another emerging technology – automated cars, (read up if you are unfamiliar with this phenomenon). I’m not going to pretend to make a claim about the development and adoption of this technology – 2013 certainly won’t be the year of the driverless car, and I don’t know which one will be – but the successes of 2012 give me an opportunity to revisit why computers are good for some things and not for others – and perhaps contrary to your intuition, computers would make much better drivers than we ever could. I am always eager to remind you that machines are not in fact very intelligent and people, despite what pop-cynicism might have told you, are really smart. The twist-ending to the story is that driving does not actually require much intelligence.
Driverless cars are one of those “magic” technologies that are either in demand or on the immediate horizon, yet I think its prospects are better than many others. As a foil to driverless cars, let us consider a competing “magic” technology: Voice recognition. From a lay perspective, both of these technologies might seem similarly complicated. In fact, voice recognition might seem simpler than driving a vehicle; after all, kids learn to communicate and talk long before we trust them with a car. The technologies face quite a different set of challenges, not all of them technical, but understanding these differences can be illuminating. I think it best to proceed by answering a string of hypothetical questions.
If voice recognition is so hard, why are we already using it? The answer is that there is a lack of risk. The biggest impediment to driverless cars, really, is that the stakes are so high. If a robot car fails, people die. When a customer service phone tree fails to understand what you are saying, it is a mild annoyance. What voice recognition we do use is unfettered with safety concerns. If you want some machine learning jargon, a problem domain includes a loss function – that is, a way to score success and failure in a way that fits the real world. The loss functions for these two problems are very different. Voice recognition that works 98% of the time is very impressive. A car that makes the right decision only 98% percent of the time is life-threateningly terrible.
Why is driving an easier problem? More pointedly, why did I say driving does not require much intelligence? People don’t want to hear the latter because driving is something they do, and it is not easy. However, people aren’t stupid because they can’t compute a square-root faster than a computer. People aren’t stupid because they can’t store a novel in their head and reproduce it word for word, (but your e-reader can).
The thing about driving is that it is a very well-defined process. For basic operations, there are clear laws, and a very well codified set of symbols that indicate all the special rules governing any location, (Stop signs, traffic lights, white dashed lines, yellow dashed lines, solid lines, lane markers, etc). For the parts of driving that require on-the-fly thinking, it is still mostly a matter of observing objects, calculating distances, “looking afew moves ahead” and avoiding collision. Driving is, in many ways, just a tedious math problem. Doing it safely requires discipline and focus and mental endurance and the ability to not let your mind wander.
If, instead of putting yourself behind a steering wheel, you put yourself behind a notebook and had to scribble out the answers to arithmetic problems as fast as you could, you wouldn’t think twice about letting a computer take over. In contrast, voice recognition suffers from all sort of inconsistent noise. Instead of the regularity of road signs and the simplicity of measuring how far away the next car is, voice recognition has to deal with the fact that people mumble, that people have accents, that they speak different languages, that they sound different in different emotional states, and that each person has a unique voice. This says nothing of the task of actually understanding language. Language is a more idiosyncratic, intuitive, /human/ thing. It’s not so easily codified.
Would a driverless car really be safer? I think so. The reason why really illuminates the difference between what computers are good at and what humans are good at. A robot car would not only know that there was a car ahead of them, but exactly how far away it was, how fast it was going, exactly how long it will take to brake to stop. A robot car would never be bored, have to sneeze, or otherwise take its eyes off the road. A robot car is looking ahead of itself and behind itself at the same time. If the robot car is common enough, it’s talking to all the other nearby cars. Now it no longer has to guess about the car ahead of it slamming its brakes – it would be told directly. The computer would be privy to information a human never would, and it would enjoy the ability to “look” in all directions at once and never be distracted. It realy is just on-the-fly number crunching, and that is not something humans are good at.
If the problem is so simple, why hasn’t it gained speed until now? How long until it is commonplace? All of the above shouldn’t trivialize how amazing the current technology is. First, a lot of what makes driverless cars possible is the ability to give a computer the same sensory facilities as a human. This is no small task, nor is it particularly affordable right now. I downplay the “intelligence” required to drive, but the AI technology needed to successfully “see” things like road signs and painted lines is definitely not trivial. The second question is much harder. I’m very optimistic about the viability of the technology, but less so about its ability to become a consumer product. The day will come, but there are two huge hurdles. One is public opinion – it seems very mixed right now. If too many people refuse to trust the technology, it will be a struggle for it to succeed. The other is legal. Even just the question of liability is enough to delay the technology for perhaps decades – when a robot car kills someone, who is at fault? Driverless cars have been legalized in three states already, but the current law holds the “driver” responsible – the car can’t be unmanned and it is assumed the driver is alert and ready to take control whenever it wants. The dream is to be able to surf the Internet or read a book or do whatever you wish while your car chauffeurs you around.
That last reality might yet be far off, but it wouldn’t be for a machine’s stupidity, I can tell you that.