DeepScale CEO Forrest Iandola Outlines His Vision for Self-Driving Vehicles
This week, Next47 confirmed it co-led a $15 million investment in DeepScale, a startup using efficient deep neural networks (DNNs) to improve the accuracy of perception systems in autonomous vehicles.
As traffic fatalities continue to claim more than 1 million lives a year worldwide, this Series A funding investment is expected to enable the expansion of DeepScale’s research and product development, supporting the company’s mission to help make self-driving vehicles and roads safer.
DeepScale is led by co-founder and CEO Forrest Iandola, a PhD graduate from the University of California Berkeley who is widely respected in scientific circles for his research around fast and efficient DNNs.
We had a chance to sit down with Forrest to learn more about DeepScale and gather his thoughts on what this latest round of investment means for his company and the future of autonomous vehicles.
Tell us a little bit about Deep Scale and what you are up to?
When I started looking into the possibilities of DNNs, I was initially focused on applying the technology to small, handheld devices. For instance, we developed a small DNN called SqueezeNet, which some scientists viewed as a gamechanger for adding deep learning capabilities to smartphones.
But very quickly, we identified an even larger opportunity to help transform the accuracy of perception systems that interpret and classify sensor data in real time for emerging, self-driving vehicles.
The key to success, we knew, would be to hold down costs since we’re talking about potentially millions of vehicles rolling off assembly lines with some form of automated driving capability by 2025. So, obviously, you couldn’t deliver DNNs as if they were going into $20,000 GPU servers that fill the trunk of a car. We believed success depended on producing fast and efficient DNNs for about $20 per sensor or processor.
That’s where we’ve been focused the last few years. We are beginning to successfully bridge that price gap, and it’s paying off with some of the investments you’re seeing in us today.
How are you differentiating yourself in the evolving autonomous vehicle market?
Most of the deep learning research community has focused on improving accuracy of DNNs without much consideration of speed or energy. But we’ve spent the last 10 years making computer vision more efficient by optimizing underlying machine- and deep-learning algorithms.
We are unusual in this space because we already understand how to make deep neural nets efficient. Instead of using a datacenter-style computing platform in the trunk of a car, we are squeezing state-of-the-art DNN accuracy onto low-cost, energy-efficient processors for the automotive market.
Today, when you open the trunk of most prototype autonomous vehicles, you find about 2 kilowatts of GPUs or tensor processing hardware worth $10,000 or more. That is more power than you need for these vehicles. It also takes up too much space, and the costs are much too high. Further, these large processing systems require fans and/or liquid cooling, which is a potential point of failure on cars that are designed to last for 20+ years.
For autonomous vehicles to become mainstream, component costs must come down, and the technology needs to be more durable. That’s where DeepScale plays. We know how to get the technology onto smaller, more proven silicon. It’s more energy efficient and affordable than current alternatives.
What obstacles do you see for autonomous vehicles, and how are you addressing them?
The regulatory environment around fully automated driving is very fragmented right now, and that’s likely to become even more true considering some of the accidents we’ve seen in prototype testing. As a result, it’s possible fully autonomous vehicles won’t be mainstream for a while.
Mindful of this, DeepScale has a diversified approach. We will continue to deliver DNN technology to enable fully autonomous vehicles, but we will also focus on enabling driver assistance features that help drivers make better decisions to avoid accidents. Government data shows 94 percent of all car accidents in the United States are a result of human error, and it’s our mission to change that.
So, on the one hand, we are delivering perception systems to help cars avoid potential collisions through automated defensive, or “guardian angel” driving capabilities. At the same time, we are enabling systems to make driving easier with features that might, for example, keep cars from straying out of their lanes.
Where do you see your company in 5 to 10 years?
Walking down the street and pointing to all kinds of cars that are safer and easier to drive because of DeepScale technology. That is the goal.
What made Next47 an attractive venture partner?
As a global venture firm affiliated with Siemens, we liked the fact that Next47 has deep roots in the automotive industry. One of the key hurdles for DeepScale has been navigating the very complicated automotive supply chain. We think Next47 will add tremendous value on the business development side, especially with respect to the large German automakers with whom Siemens has strong relationships.
When you boil it down, Next47 has several things that we were looking for in a strategic partner: great connections with a global ecosystem of customers and partners, abundant financial resources with its $1.2 billion global venture fund, and a team that is genuine and smart.