How do self-driving cars work? Do they actually "see" the world around them?
The short answer: No.
I'm just a layperson trying to understand processes that I don't hear many experts trying to explain to other laypeople. But as far as I can understand, self-driving cars work through a combination of software and sensors.
What part of it involves AI?
Well, using, in part, deep neural networks is how these systems are trained to "see." This training involves a host of low wage workers around the world, many working for third party companies, signing NDAs so word doesn't get out about what it is that they're doing. Most of them don't even know. And, as I learned from a chapter in Madhumita Murgia's brilliant CODE DEPENDENT, these workers will be sent clips of a camera trained on the road. They will then spend hours and hours (days and days, months and months, you get the point) marking objects. Identifying them on screen. Cataloging them. Labelling them (often a process also involving online translation systems). This data is then used to train systems to identify objects (oversimplified, but they are finding patterns).
This is all to say, there is no "seeing" going on, not in the human sense. And it would be very helpful if more people who understood these processes broke them down and explained them to the public. It might help dispel some of these "artificial intelligence" myths. But, of course, if AI loses its magic, maybe we would be a bit more hesitant to rely on it across the board without intensive testing and investigations first. Without, well, critical thought.
Ask the neuroscientists in the room: how much do we even know about how the human brain works? Do you think anyone has figured out how to replicate something that is still, even with all of the progress made, pretty much unknown to us?
One final point: now is a crucial time for us to support journalists who are working to understand and help us all understand and better talk about how these systems work. To help us dispel myths, as Murgia, for example, so keenly does.
Some excerpts (published on LitHub, link below) from her book:
"Today, real-world AI is less autonomous and more an assistive technology. Since about 2009, a boom in technical advancements has been fueled by the voluminous data generated from our intensive use of connected devices and the internet, as well as the growing power of silicon chips. In particular, this has led to the rise of a subtype of AI known as machine learning, and its descendent deep learning, methods of teaching computer software to spot statistical correlations in enormous pools of data—be they words, images, code or numbers.
One way to spot patterns is to show AI models millions of labelled examples. This method requires humans to painstakingly label all this data so they can be analyzed by computers. Without them, the algorithms that underpin self-driving cars or facial recognition remain blind. They cannot learn patterns."