Computer vision and machine learning advancements are supporting a variety of new functions within vehicles — including the automatic detection of pedestrians and objects.

Automotive neural networks, for example, can “train” a computer to infer rules based on learned examples.

But how do you regulate a system that, in effect, is learning as it goes?

In a webinar this week titled Automated and Connected Commercial Vehicles: Next Steps for ‘Hands-Off’ Operation, a reader had the following question for two presenters:

"What are the regulatory hurdles related to vehicles and machine learning?"

Chris Woodward from off-highway hydraulics manufacturer Danfoss Power Solutions, as well as Jace Allen from the automotive simulation provider dSPACE, shared their thoughts. See their responses below.

Chris Woodward, Business Development Manager of Autonomous Machines, Danfoss Power Solutions: When it comes to autonomous vehicles, we could probably spend all day talking about the regulatory side of the world. I’m not a functional safety consultant myself, but I can weigh in on what I’ve seen.

In our investigation, the European certification bodies haven’t been very comfortable with the sort of “black box” that the neural network is on the machine. They’re used to seeing and certifying more hard-coded algorithms that they can run the test cases against and validate.

Proving out the safety and reliability of a neural net becomes a bit more complex. In the current state, things are just a little more uneasy. I don’t think these are challenges that can’t be overcome, but as it exists today, the technology is moving faster than the regulatory bodies, especially in Europe. That’s really the challenge I see there.

Jace Allen, Business Development Manager of Simulation, Test, and Electrical/Electronics Data Management, dSPACE Inc.: That was always kind of a scary thing to me, that I potentially have this learning system that executes in the vehicle.

The thing that you have to take into account here is that the learning has been done in a system like this. When a system is deployed into a vehicle, it’s really not in a learning mode anymore. It has been trained and has been implemented as a trained system. It has to go through the same type of testing that has been done on any type of embedded system — in terms of having to pass a series of tests and having to be validated and certified at the level it was done at.

There are potential areas that need to be considered, like on-air updates for example, but, once again, they’re going to have to go through regulatory hurdles in order to ensure robustness in the system that has been implemented or upgraded.

Chris brought up a good point. Europe is kind of ahead of the curve in terms of the NCAP processes. There’s a lot of work being done today by various groups in the U.S., including SAE and NHTSA , to try to bring together these standards, find a validation process, and try to certify them as an approach that people need to follow in order to get to homologation of the system.

What do you think about machine learning and neural networks in vehicles? Share your comments and questions below.