There are a lot of potential problems with the way we train AI models. The datasets used to teach machines to think have been repeatedly shown to create biased algorithms, sometimes leading to discrimination. By some estimates, data scientists spend 80% of their valuable time acquiring, cleaning and managing datasets. And in some cases, datasets don’t even exist in the first place.
Physicist Nathan Kundtz has a plan to fix that, which could drive huge improvements to AI training.
Kundtz, who previously led Bill Gates-backed satellite communications company Kymeta as CEO, has launched a company called Rendered.ai to create synthetic datasets for training artificial intelligence systems, as well as building an infrastructure that allows engineers to customize and modify them.
The Bellevue, Wash.-based startup is currently working on AI that learns from visual datasets, including images and videos, as well as data from radar, X-rays and thermal imaging. This week it announced a $6 million seed round.
There are other companies providing synthetic datasets for AI training, creating images for machines to learn from when they can’t be collected in real life. But they take a different approach.
“Most of them think about this problem as a computer graphics problem. So the product that they provide and the solution to that is a set of images,” Kundtz said.
“We think about that very differently. We think about this as an engineering problem. And the product that we provide is a platform that gets you to business outcomes necessary to drive your business forward.”
Kundtz provided an example. There are researchers who monitor economic activity by using satellite images. They use AI to scan the images for construction cranes, trucks and other physical manifestations of trade and growth. But before that can happen, the model might need a training dataset with millions of pictures to capture examples of these objects, plus the skills of an expert annotator to recognize and identify them. And then there are irregularities to contend with — maybe snow covering a crane, cloud interference or rust that obscures it — causing an annotator to miss the item.
To solve that problem, Rendered creates its own 3D mini universe and populates it with the cranes and trucks for the AI to learn from. They can add the same irregularities, but still find the cranes and trucks. It’s a solution that might sound just as laborious, but the simulation relies on basic science to carry the load.
“Especially in the computer vision world, so much of the data is really based on physics that we understand really well — like light interacting with things is [an area] where we have a tremendous experience and incredible simulation tools,” Kundtz said.
Rendered puts that dataset into an infrastructure so that AI engineers can modify and customize it.
It’s an ambitious vision, but Kundtz brings entrepreneurial credibility to the venture.
Kundtz began his career as a manager at Intellectual Ventures, the high-tech innovation hub and brainchild of former Microsoft Chief Technology Officer Nathan Myhrvold. While there, Kundtz led a team working on cutting-edge antenna technology, which became the foundation for the spinoff startup Kymeta. He was Kymeta’s CTO and then CEO before leaving in 2018.
Kymeta has raised more than $440 million, according to PitchBook — with roughly half of that accrued under Kundtz’s leadership. The company continues to rack up milestones with its next-generation satellite broadband technology.
It was the experience at Kymeta that helped inspire Kundtz’s newest venture. When interacting with satellite companies, he saw that their work was sometimes limited by a lack of datasets for training AI.
In addition to satellite companies, Rendered’s platform can provide datasets for engineers in medicine, transportation, robotics and security. The startup is currently in beta mode, working with a handful of users to fine tune its technology. The seed funding, which came from Tectonic Ventures, Congruent Ventures, Union Labs and Uncorrelated Ventures, will help the company expand its services to more customers.
Kundtz is excited about the work that he has embarked on, addressing what he sees as a fundamental hurdle for AI.
“The problem is a little bit of a ‘well duh’ problem,” Kundtz said. “And yet when you see something that so foundational and yet not being addressed in a way that is scalable and really getting to the heart of the problem, that’s the kind of thing that gets me up in the morning.”
