Volvo Cars Autonomous driving
​Image: Volvo Cars

Can we trust a self-learning machine?

​Security and trust is needed before self-driving cars can finally be launched. With the help of machine learning and deep learning, key technologies are being developed that will take us there – the self-learning system.
A car approaches a crossing where a pedestrian stands next to the road, turning her head. Is she about to cross the street or not?
“People react quickly, almost without thinking, but it has been difficult to get a self-driving car with traditional computer vision technology to extract such information. With deep learning, it becomes possible”, says Erik Rosén, technical specialist at Zenuity.
 
The company Zenuity, co-owned by Volvo Cars and Autoliv, will develop software to ensure the safety of self-driving cars. It's about getting the car, or as Erik Rosén says, "the four-wheel computer", to find patterns and relations in large amounts of traffic data, and thereby gaining the ability to drive safely in traffic.
“Deep learning, which is machine learning using deep neural networks, has become hot in recent years. The technology is invaluable to us, especially when we develop what we call the perception layer.”
 
In order to gain excellence, Zenuity works closely with Chalmers. In the division of Computing Science they have several industrial PhD students, and Nasser Mohammadiha, technical specialist in machine learning at Zenuity, is an Adjunct Associate Professor at Chalmers.
"In my dual role, I also give researchers insight into the challenges facing Zenuity, which gives rise to new research questions. Chalmers is an important part of this development.”
 
Several research projects have started in collaboration, including verification of software.
"We want to understand why the system makes a certain decision. How does it happen? To do that we have to go back to the origin of the source. We make tests with both good and bad decisions. A safe system must be able to anticipate all possible scenarios, such as how other cars are expected to drive. It's very complex”, says Nasser Mohammadiha.
 
The development is extremely rapid and companies are competing to get a self-driving system on the market. Are they in too much of a hurry? Erik Rosén is afraid so.
"Some technology companies are very keen on performing cool demonstrations. They show off a powerful super sensor, which is great at capturing information. But what happens if it stops working, if it’s not able to connect to the map or anything breaks?”
 
At Zenuity, they work with the motto "make it real". A self-driving system might be really smart in traffic, but as long as it cannot be guaranteed that it will always work – it's not safe. And the software architecture is not yet there.
"There is a need for redundancy in the software architecture as well as in the hardware. That is the biggest challenge. If a computer stops working in the car, another must be switched on, if the sun is low and blinds a camera, a radar or other sensor must take over.”
 
It sounds almost like Nasa technology?
"Yes, actually," says Erik Rosén, laughing. Certainly, if you want to take the control of the car from the driver, you cross a line that is very challenging. Anyone who wants to put a product on the market must work with redundancy", says Erik Rosén.
 
He says that Sweden is at the forefront of the development of self-driving systems, and already in 2021 he hopes that Zenuity has one in the market.
“Which means that the driver can hand over the control to the car but only under certain circumstances. I cannot say exactly which today, but weather conditions, lighting conditions and traffic environments are what matters", says Erik Rosén.
 
 
 
  • Machine learning is simply put, algorithms that are trained to solve tasks based on recognizing statistical patterns in large amounts of data.
  • Deep learning is machine learning that uses so-called neural networks (see below) as a model for learning.
  • Artificial neural networks are self-learning algorithms that imitate the model of biological neural networks (such as the brain). Artificial neural networks can often handle problems that are difficult to solve with conventional task-specific programming. A neural network must be trained with examples before it can fulfill its intended function.
 
 

Welcome to our Initiative seminar on Digitalisation:

Security & Privacy | Machine Intelligence

On 15 March 2018, Chalmers organise a second Initiative seminar on Digitalisation. This time we present a more in-depth programme – with half a day on Security and Privacy and the other half-day on Machine Intelligence. 

Published: Mon 26 Feb 2018.