Continental presented plans recently to investors at Bank of America in New York: the supplier wants to work with Autobrains (formerly known as Cartica AI), a startup from Israel applying a new version of artificial intelligence called “unsupervised AI”. This new approach, they say, could give artificial intelligence for self-driving cars the decisive impetus.
In autonomous cars, AI derives and executes the correct driving maneuvers from the data from sensors such as cameras, radar, ultrasound, and lidar. The biggest challenge so far has been the enormous effort that is required to train the AI. For example, image recognition algorithms, of central importance in autonomous driving, have to be trained with billions of images in countless repetitive loops until they clearly identify certain objects or living beings. Pictures of children playing and of stop signs, for instance, are shown to the AI system until the system recognizes them with 99.9999 per cent (6 Sigma) reliably. In difficult borderline cases, such as pictures with graffiti, in bad weather conditions, at dusk, and in chaotic environments, enormous amounts of data are needed—which costs a lot and takes time.
The new approach has been researched for a long time and works under the generic name “non-guided AI”. In contrast to the conventional approach, the AI programs should develop reliable criteria for object recognition themselves and refine their own algorithms. According to the Continental presentation, development of the new AI systems gets by with a tenth of the previous amount of data and computing power. This would significantly reduce development times and costs.
Former Continental boss Karl Thomas Neumann sits on the Israeli company’s supervisory board. “Unsupervised AI is extremely exciting because it calls into question the entire mainstream of current AI developments relating to autonomous driving,” he says. If the concept catches on, self-driving cars will be more quickly deployable because they’ll be better able to cope with new situations for which they are not yet explicitly trained.
