Can We Train Self-Driving AI to be 100,000x Cheaper?
Annotation is an expensive bottleneck. There’s a workaround.
“How does your AI drive on a road it’s never seen before?”
That’s the first question posed to Helm.ai CEO Vlad Voroninski by TechFirst podcast host John Koetsier during Vlad’s recent guest appearance.
It’s a great question, of course, and offered us a unique springboard into talking about other opportunities for our proprietary Deep Teaching technologies for Unsupervised Learning.
The answer is almost better rephrased as a question: “Well, how do you drive on a road you’ve never seen before?”
You learned how to drive, but that’s not all. First, you learned how to see, and then how to recognize and interpret the things that appear in your field of vision, and then you learned how to make the mechanical movements to react to those objects.
In your lifetime, you solved a huge number of problems at human levels of accuracy. Most of that learning, with some early or occasional instruction, was unsupervised. Through your own observation, research and experience, you learned to drive … but not in the 12 or so hours you spent in Driver’s Ed class.
Most of the education you received occurred in the lifetime you spent seeing, learning how to recognize objects, finding your way to places, watching your parents drive, or driving since getting your permit. You mostly taught yourself.
That’s what Deep Teaching is: a way of teaching AI software to learn to drive the way you learned (and are still learning) to drive. And, as our CEO says in the interview, “Humans … when they’re paying attention, and not inebriated or angry, are very good at driving.”
Humans still get into six million auto accidents per year, though. So, if we were merely aiming for human-level accuracy, that wouldn’t be good enough to launch a fleet of self-driving cars. It’s important to train AI to be much safer than humans on the road.
For years, the thought was to annotate images — a function often performed by humans — to give the AI training examples to learn. That’s an expensive endeavor, and that cost is a bottleneck for the autonomous vehicle industry. What if we could come up with a learning technology that requires no annotation? What if the computer could “see” and learn for itself?.
“The software we develop has much higher levels of accuracy. It can generalize better and handles more corner-cases. Using computer vision allows you to build an ultimately cheaper [AI] stack,” says our CEO.
“We can drive with just one camera and one GPU. We can drive it on a road without using data from that road. The difference is the ability to interpret real-time sensor data quickly and accurately.”
But how do we generalize better, and interpret real-time sensor data quickly and accurately? That’s the essence of Deep Teaching, our proprietary technology steeped in applied mathematics and, as Mr. Koetsier alluded, a bit analogous to Gestalt principles.
We likened it to solving reconstruction problems in compressive sensing. AI observes information about an object of interest, then, by incorporating structural and relational assumptions about that object of interest, processes it in a way similar to how MRI technology’s been sped up by a factor of 10, but MRI with Deep Learning speeds it up by a factor of 100.
As it pertains to autonomous vehicles: The AI teaches itself what it needs to know. It knows the world is 3D, it can associate learnings by temporal and spatial contiguity, and it can address and ascertain object permanence, by performing semantic segmentation tasks and predicting what every single pixel in an image is — or what it corresponds to. We demonstrated the results at CES this year.
Humans mapped the human genome in similar fashion. It was costly and arduous at first. Yet over time, the cost of mapping the human genome has “dropped 100000x over the years.”
There’s definitely the potential for this to happen in AI over the next few years. That’s what we’re working towards: paving the way to train self-driving AI at a minuscule fraction of the current cost.
The challenge, as said on the podcast, “is not just removing the need for annotation but making sure there’s enough learned per-image … The sophistication of your AI stack will determine if these technologies can be used at scale.”
When that happens, there’s a lot of applications for Deep Teaching. Autonomous driving is a problem we’re confident we can solve with it, yet there are plenty of possibilities in potential trillion-dollar industries, like delivery robots on land and air, as autonomy comes to full fruition.
Deep Teaching is the unsupervised future of AI. We’re excited to learn what happens next. If you’d like to learn what else we’re excited about, check out the full interview here:
Huge thank you to TechFirst and John Koetsier for having us on the show.
Interested in learning more about Helm.ai’s approach to Unsupervised Learning and AI for autonomous driving? Be sure to subscribe to the Helm.ai YouTube channel for more context or visit the Helm.ai website.