A Waterloo startup has partnered with a German automotive giant to demonstrate how its artificial intelligence technology can potentially accelerate the development of the electronic brain behind autonomous vehicles.
Artificial neural networks simulate the human brain's ability to make decisions, to learn and to adapt to the environment. A team of engineers at Audi saw a reduction of more than 90 per cent in the number of hours spent processing and refining data for those networks using technology developed by Waterloo-based DarwinAI.
A paper outlining their work was presented this week at the British Machine Vision Conference at Cardiff University in Wales. It shows how the automaker used the local tech company's expertise to develop faster and more compact neural networks for self-driving cars, without sacrificing accuracy.
"There were real productivity benefits they found for their developers," DarwinAI CEO Sheldon Fernandez said in an interview with The Record. "It's exciting for us ... we're finally able to publicly show the validation, and that's really important to us as a young startup."
The German auto manufacturer was one of the tech company's first clients after DarwinAI announced itself publicly last year.
The research paper studied the impact of the company's Generative Synthesis (or GenSynth) technology to design neural networks for object detection in autonomous vehicles.
It found that the design platform "can not only produce better and more compact networks tailored for specific tasks more quickly but also help designers gain better insights and improve their own ability to design deep neural networks."
Neural networks were created for 10 classes of objects, including buses, cars, trucks, fire hydrants, bicycles and traffic lights. Researchers found they could achieve the same or better average precision with the GenSynth technology as they did with a network that was designed by humans, and it was also more compact and had lower computational costs.
The reduction in cost and size is critical as neural networks have traditionally been too large for consumer electronics.
Overall, the number of hours programmers spent refining models was reduced from 200 to 17, and the number of hours spent processing the data dropped from 10,400 to just 760.[Read more]
(This article was originally published on therecord)