Going beyond 10 million miles
If you’re building the future, it’s hard to avoid a little bit of hype. The promise of a better future inspires entrepreneurs, investors, and consumers alike who keep the cycle of innovation moving. But hype can be tricky, especially for AI. Understanding what AI can do, and should do, is tricky even for experts.
The autonomous car arms race is in full force with the world’s biggest players jockeying for position both technologically and public opinion-wise. Currently, 60 companies are authorized to test their self-driving cars on the road in California, and every few weeks, our industry is in the news with a new timeline, demo, or product launch.
Often, the arms race is quantified in miles driven, which makes some sense. Statistics on total miles driven are an essential proxy for how seriously players are about reaching Level 4 — highly automated driving — and beyond. But these statistics are also part of the hype. They’re used for marketing purposes or to intimidate rivals into believing that one player has an insurmountable lead over others in bringing AVs to market.
For example, Waymo announced that its cars hit 10 million miles last month and this week, they were given the green light to test fully driverless cars without safety drivers in California. Uber, Tesla, Mobileye, Cruise, and others are not far behind — all with the goal of reducing traffic accidents. Yet, despite our cars driving more miles, we’re still creating accident-prone vehicles.
Since the California DMV instilled new regulations on self-driving cars in 2014, it has collected more than 100 AV collision reports, half of which happened this year as more cars hit the streets. This begs the question about whether there’s a deeper story than “more miles = better cars.” Are there situations where no amount of driving is sufficient?
There’s no doubt that autonomous cars need to drive on real roads. Machine learning algorithms gain accuracy with more data, and for self-driving cars miles equals data. You can test the vehicles in simulation — which is handy, because the most important data can come from circumstances that don’t happen very often. Data only helps when it contains the information you need to know. But we can’t even imagine how much information we need when driving cars because, as humans, we rely on a vast, innate ability to make sense of the world around us — something that you don’t learn just from driving more miles.
When a human — the reigning champion at the task of driving, so far — drives a car, what are they doing? On which skills and knowledge do they depend? How did they learn those skills? When thinking about these questions, we begin to see how driving might involve more than just driving.
The subconsciousness of driving
The human brain is good at telling stories about the world. When asked what skills we use to drive, we have plenty of answers. You need to be able to steer and know where the edges of the car are. You need to know the rules of the road and what different signs mean. You need to be able to avoid obstacles, match the speed of traffic, and know where the road is and on what part of it to drive.
Granted, everything on that list is necessary for driving, and they tell a pretty good story of what driving entails. The one common characteristic of all of those skills is that they are things that a human had to learn to pass a driver’s test. However, they lack the elements of driving that are so effortless that we don’t even notice doing them — and so automatic that even if asked how we did them, we wouldn’t know how to answer.
An example that isn’t directly related to driving is facial recognition. Imagine the face of your best friend. Now, explain how you’re able to recognize them. You might begin by listing their facial features or the color of their eyes and hair, but many people fall under the category of brown hair and brown eyes.
So then, your brain might tell a story with other reasoning like, “he always wears that hat,” “she has her mother’s nose,” or “I’d know that hair anywhere.” The underlying answer, really, is “they look like themselves.” You don’t have conscious access to the features that you use to recognize a familiar face. The features are there, and you rely on them, but you don’t necessarily have the ability to describe them.
We have evolved to be good at doing something without conscious access to how we do it. These abilities aren’t learned skills, and for that reason, they tend to be the hardest things to teach computers. Our ability to understand other humans and predict their actions falls in this category. We have evolved to do it. The ability to make cooperative social judgments might be the single evolutionary change that led to the explosion in human intelligence.
The combinatorial explosion in the mirror
Looking at somebody else and making an inference about what’s in their head turns out to be something that we do a lot when driving. When we see somebody on the street, whether they’re in a car, on a bike, or walking, we make inferences about them. These inferences may or may not directly correlate to the task of driving but are vital for driving cars.
Take understanding whether somebody wants to cross the street. Someone could be standing next to the road, but maybe they’re waiting for a bus. They could be walking into the road, but getting into a car. They could be standing still waiting for a gap in traffic so that they can cross or walking parallel to the road for the same reason.
Humans can judge these circumstances instantly and effortlessly, while machines rely on contextual cues — the presence of a crosswalk or a stoplight — that are informative but insufficient. To understand what somebody’s going to do from context, not only do you need all the context — that speeding car around the corner that the pedestrian can see, but you can’t — but you need many, many examples of it.
The amount of data needed is subject to something mathematicians call a “combinatorial explosion.” The further you’re trying to predict, the harder the prediction gets. Even if we drive a billion miles, we still won’t have seen all the combinations of factors that could influence somebody stepping in front of our car.
Human behavior is one of the most complex problems of all. It’s not even clear how much data we may need to understand humans — or whether we’d be able to collect enough data to teach a machine how to make judgment calls. We don’t learn these things; we are born with the ability to acquire an incredibly sophisticated set of social reasoning tools subconsciously.
As the dust begins to settle on the hype of the promise of self-driving cars, we’re increasingly aware of just how complicated it is to teach a machine how to drive. We have the driver’s manual, and we’ve taught computers how to follow them, but we need to go beyond the black and white — beyond the miles driven — and begin thinking more critically about the cognitive science behind the wheel.
