字幕表 動画を再生する 英語字幕をプリント SciShow is supported by Brilliant.org. By now, you've probably heard that self-driving cars are coming soon. If you haven't—surprise! They're coming soon! But people have been saying that for at least a decade, and I still can't buy a car that'll drive me to work while I nap in the passenger seat. Some cars already come with partial autonomy, systems like Tesla's Autopilot, that assist drivers or sometimes even take control. But they still need a human driver who can grab the reins on short notice if things get dicey, which is why someone in the UK got arrested earlier this year for trying the passenger seat thing. There are some fully driverless vehicles that might be released in the next few years, but they're only meant for very specific uses, like long-haul trucking or taxis confined to certain streets and neighborhoods. That's because general-purpose driving is hard! The software has to work out a lot of really tricky questions to turn information from its sensors into commands to the steering and pedals. And despite all the money and brainpower that's being poured into research, there are still major challenges at every step along that path. The first thing a self-driving car has to do is figure out what's around it, and where everything is. It's called the perception stage. Humans can do this at a glance, but a car needs a whole cornucopia of sensor data: cameras, radar, ultrasonic sensors, and lidar, which is basically detailed 3D radar that uses lasers instead of radio. Today's autonomous vehicles do pretty well at interpreting all that data to get a 3D digital model of their surroundings the lanes, cars, traffic lights, and so on. But it's not always easy to figure out what's what. For example, if lots of objects are close together say, in a big crowd of people it's hard for the software to separate them. So to work properly in pedestrian-packed areas like major cities, the car might have to consider not just the current image but the past few milliseconds of context, too. That way, it can group a smaller blob of points moving together into a distinct pedestrian about to step into the street. Also, some things are just inherently hard for computers to identify: a drifting plastic bag looks just as solid to the sensors as a heavier, and more dangerous, bag full of trash. That particular mix-up would just lead to unnecessary braking, but mistaken identities can be fatal: in a deadly Tesla crash in 2016, the Autopilot cameras mistook the side of a truck for washed-out sky. You also need to make sure the system is dependable, even if there are surprises. If a camera goes haywire, for example, the car has to be able to fall back on overlapping sources of information. It also needs enough experience to learn about dead skunks, conference bikes, backhoes sliding off trucks, and all the other weird situations that might show up on the road. Academics often resort to running simulations in Grand Theft Auto yes, that Grand Theft Auto. Some companies have more sophisticated simulators, but even those are limited by the designers' imaginations. So there are still some cases where perception is tricky. The really stubborn problems, though, come with the next stage: prediction. It's not enough to know where the pedestrians and other drivers are right now the car has to predict where they're going next before it can move on to stage 3: planning its own moves. Sometimes prediction is straightforward: a car's right blinker suggests it's about to merge right. That's where planning is easy. But sometimes computers just don't get their human overlords. Say an oncoming car slows down and flashes its lights as you wait for a left. It's probably safe to turn, but that's a subtle thing for a computer to realize. What makes prediction really complicated, though, is that the safety of the turn isn't something you just recognize it's a negotiation. If you edge forward like you're about to make the left, the other driver will react. So there's this feedback loop between prediction and planning. In fact, researchers have found that when you're merging onto the highway, if you don't rely on other people to react to you, you might never be able to proceed safely. So if a self-driving car isn't assertive enough, it can get stuck: all actions seem too unsafe, and you have yourself what researchers call the “freezing robot problem.” Which itself can be unsafe! There are two main ways programmers try to work around all this. One option is to have the car think of everyone else's actions as dependent on its own. But that can lead to overly aggressive behavior, which is also dangerous. People who drive that way are the ones who end up swerving all over the highway trying to weave between the cars. Don't do that, by the way. Another option is to have the car predict everyone's actions collectively, treating itself as just one more car interacting like all the rest, and then do whatever fits the situation best. The problem with that approach is that you have to oversimplify things to decide quickly. Finding a better solution to prediction and planning is one of the biggest unsolved problems in autonomous driving. So between identifying what's around them, interpreting what other drivers will do, and figuring out how to respond, there are a lot of scenarios self-driving cars aren't totally prepared for yet. That doesn't mean driverless cars won't hit some roads soon. There are plenty of more straightforward situations where you just don't encounter these types of problems. But as for self-driving cars that can go anywhere… let's just say the engineers won't be out of a job any time soon. I love the layers of thinking involved in this kind of problem solving. And while I'm not an engineer designing self-driving cars, but I still get to practice this kind of thinking on Brilliant.org. Right now, I'm working through the Convolutional Neural Networks lesson to help me learn how to work with neural networks. I've already gone through the overview, and this “Applications and Performance” quiz has a car on it, so that's what I'm going to try my hand at next. The quiz already explained how this network works. And then it's asking how we should modify it to suit this imagenet challenge, to help it categorize objects better. I think the answer is C: to add a fully connected network at the end to help predict probabilities for what the object is, based on the high level filter activations. And I got it right! What's great about these quizzes is that they keep building on each other, so even though I got that one right, I'm still getting more information throughout and each question gets a little bit more interesting. And if you get one wrong, that's ok too! Because the point isn't to beat the quiz, it's to keep learning, just like these neural networks do! So, if you want to test out YOUR neural network, the first 200 viewers to sign up at brilliant.org/scishow will get 20% off their annual premium subscription, and you'll help support SciShow - so thanks!