White flakes float down, the road becomes blanketed in snow and becomes slippery. What do you do as a driver? Drive carefully, steer, brake – sure. But how should an autonomous vehicle behave in variable driving conditions? How to determine the slipperiness of the road, how to calculate the complex interaction of all factors? “What is simple for machines is usually difficult for people – and vice versa," says Paul Hochrein, who researches autonomous driving at Volkswagen Group’s driving dynamics control division.
Research cooperation between Volkswagen and Stanford
The artificial intelligence (AI) of an autonomous vehicle has to recalculate the process that usually happens unconsciously for humans several times per second, compare it with internal data records and the input of numerous sensors, keep an eye on the environment, follow numerous written and unwritten rules and react to new situations in a fraction of a second. Among other things: recognizing how slippery the road is, and driving accordingly. This is an important area of autonomous driving in which Volkswagen Group is conducting research together with Stanford, California’s elite university.
Extreme situations during autonomous driving
The focus is on the application of artificial intelligence in autonomous and assisted driving in all conceivable situations on the road. “We want to prove to the car that this is possible,” says Hochrein, referring to the renowned Pikes Peak race in Colorado, in which the Volkswagen Group has already set various records. In general, there are two main challenges for AI in autonomous driving: driving at the limit on the racetrack and driving in city traffic. “If a complete system including AI can handle such extreme situations, you’re already a huge step ahead,” says Hochrein. Of course, the corresponding sensors and data processing must also be available as a basis for this.
Whether on a racetrack or in a big city, how should an autonomous vehicle deal with snowfall and other road conditions? Volkswagen Group and Stanford have come up with an answer to this question: an artificial neural network that can recognize different surfaces and react accordingly. And it’s better and more flexible than a classic program that can only perform tasks such as semi-autonomous driving with immense prior programming effort. Artificial neural networks are a special form of artificial intelligence – and of great importance for the development of autonomous and assisted driving.
Not accurate enough for the real world
Hochrein explains artificial intelligence as follows: “Mathematical functions that act as a human would.” But there is a problem with autonomous driving. “If we use simple algorithms, they do what they are supposed to do. But they do not represent reality accurately enough for safe autonomous driving," says Hochrein. After all, enormous amounts of different data have to be processed when driving: from the steering wheel angle to the possible behavior of an oncoming motorcycle, to the question of whether the traffic lights in front are red or green and whether they are aimed at the driver’s own lane or not.
“To represent all of this mathematically, leads to very complex formulas,” says Hochrein. This is why something different is needed here – a program that can independently deal with large amounts of data and derive a certain behavior from it. A kind of mechanical brain, consisting of digital ones and zeros, that does not have to be programmed down to the last detail, but rather programs itself with human help in such a way that the desired result is achieved in the end.
100 trillion synapses
100 trillion synapsesThe physical blueprint, the human brain, consists of up to 100 billion nerve cells, also called neurons. And each of these neurons in turn maintains around 1,000 connections to the other neurons around it – sometimes more, sometimes less. These connections are called synapses, so there are about 100 trillion of them per brain. In its written form: 100,000,000,000,000,000 pieces. They form a highly complex network that fulfils numerous functions – and that inspires not only brain researchers, psychologists and physicians, but also computer scientists.
This is how it works in the human brain: chemical and electrical impulses rush through the synapses from neuron-to-neuron, which are then activated in certain patterns, i.e. “fire”. Synapses and neurons fire in certain patterns and rhythms. Learning processes in the brain can be imagined in such a way that certain neurons are activated more frequently than others, whereby some synapses become stronger, so that beaten paths become highways – and vice versa: synapses that are used less often break down.
Experimenting with dog photos
The same applies to artificial neuronal networks: They try to find out which patterns produce the desired results. The network itself selects which neuron is linked to which other neurons and to what extent. In other words: The original programming only provides the rough framework – and AI fills the details independently on the basis of specifications. These patterns are retained and reinforced. However, this is only possible if it is trained correctly.
An example: A neural network should learn to recognize photos of dogs – and not to confuse them with those of cats, for example. It receives different sets of photos of different dogs, as well as of cats, and the corresponding information as to which output is the right one.
Artificial neural networks and deep learning
Artificial neural networks have to process a lot of information at once. This requires parallel computers. It’s not CPUs that are used for this purpose, but graphics cards, GPUs that are connected in parallel and that were developed primarily for the gaming industry. “Without them, we wouldn’t be where we are now,” says Hochrein. Depending on the type of network, there are different layers between the input and the output. The more layers, the more computing power, the more artificial neurons and synapses, the better the performance. Overall, it is a system of mathematical equations, with flexible parameters and a large number of different possible combinations between the different neurons – between 400 and 40,000,000. The more layers, the deeper: this is Deep Learning, a subtype of artificial neural networks.
At the beginning the program might still make many mistakes – there are many different dog breeds, with different forms of ears, muzzles, body sizes. It independently analyses and experiments the numerous pixels, recognizes patterns and distinguishing features, and correlates them. And it learns which parameters it has to change in such a way that the way through its innermost, from the input of the photo to the output “This is a dog” is correct.
Are Chihuahuas dogs?
At some point the program will be able to determine with a certain probability that a Chihuahua counts as a dog, without it having been explicitly confirmed beforehand as the correct solution. Depending on the technical equipment and programming, this can take different lengths of time. How it creates the connection from photo to result and which neurons fire, is left to the network itself with its numerous nodes and connections. The main thing is that it works. Perhaps the best-known example of a recent artificial neural network is AlphaGo, a program from DeepMind, belonging to Google.
AlphaGo is a world champion
For a long time, the Asian board game Go was considered one of the last bastions of human intelligence and intuition. It is far more complex than chess: it is said that there are more possible combinations in one game than there are atoms in the universe. But this fortress was also stormed in 2016: a neural network was trained for months using Deep Learning, a special type of neural network that can learn particularly well on its own. It played more than 30 million games. And then defeated the previous Go World Champion, Lee Sedol from South Korea, who claimed the number one spot for a decade.
“That was the breakthrough,” says Hochrein. “The basic idea for artificial neural networks had been around for a long time, but now advances in computer science have combined with advances in computing power. As far as autonomous driving is concerned, Hochrein is optimistic: “Level 5, i.e. fully automated driving, is basically feasible, but it is a highly complex task to get it safely on the road.” You shouldn’t just rely on neural networks – classic programs are also suitable for certain tasks. “Neural networks are not a remedy for everything,” says Hochrein, “but they are enormously important for the development of autonomous and assisted driving.”
1 Concept car