More Deep Learning Experts Needed?

Article By : Junko Yoshida, EE Times

No straight path to deep learning enlightenment, says DeepScale founder

MADISON, Wis.— DeepScale, a Mountain View, Calif.-based startup founded in 2015 to develop deep-learning perception software for ADAS and self-driving vehicles, has just closed its Series A funding with $15 million from two venture firms, Point72 and next47.

DeepScale's deep neural network software detects vehicles, pedestrians and objects of significance for automated driving, using low-power automotive-grade chips. (Source: DeepScale)

DeepScale’s deep neural network software detects vehicles, pedestrians and objects of significance for automated driving, using low-power automotive-grade chips. (Source: DeepScale)

The startup also unveiled a new partnership with Hella-Aglaia Mobile Vision GmbH. Earlier this year, Visteon, a tier one based in Michigan, announced that it has partnered with DeepScale as Visteon developed its first autonomous driving technology platform, DriveCore.

Most revealing, however, in a phone interview with DeepScale CEO Forrest Iandola, was his acknowledgement that the world doesn’t have enough deep-learning experts.

Not enough deep learning experts
Both car OEMs and tier ones’ appetite for software expertise — deep learning in particular — has only grown over the last 18 months. The industry, in general, suffers a chronic knowledge gap in deep learning and how to leverage it to develop software.

Even DeepScale, co-founded by Iandola, a PhD from U.C. Berkeley working on deep neural networks and computer vision systems, feels pressed to internally scale up its expertise more quickly to meet the external demands.

Asked how DeepScale intends to use the $15 million, Iandola told EE Times, “We not only need to hire deep learning experts but to develop an internal [deep learning] training program to scale the team.”

In essence, the field of deep learning is still in its infancy — even in academia, acknowledged Iandola. So, scaling the team within the startup won’t be as easy as just hiring a bunch of people with PhDs in deep learning.

Forrest Iandola

Forrest Iandola

Calling deep learning “pretty much an interdisciplinary field,” the CEO of DeepScale told us, “You need not just experts in math and algorithms, but also computer system experts who know how to implement software on the platform. Then, we need experts familiar with data pipeline who can align and curate data set.”

In short, deep learning experts are people whose knowledge matches up in unprecedented ways. It requires specialists in deep learning models, plus people who understand deep learning infrastructure, and others experienced with deep learning data sets. Iandola said, “We plan to offer internal classes and assigning a mentor to every two to three more junior engineers.”

The goal is to bring up internal software engineers to become versed in deep learning in one to two years, he explained. Asked how internal training at DeepScale differs from the graduate courses that teach deep learning, Iandola said, “It will be similar. We will replicate what’s been taught at Berkeley.”

However, that said, “It’s not like we have a deep learning textbook. We will be teaching problem-based learning,” Iandola said. When deep learning experts are faced with real-world problems in autonomous driving, for example, they have a lot to share with the rest of the team.

DeepScale started this year with 12 people. With the latest hires, “We are a company with 18 people. Most of them are engineers,” said Iandola.

DeepScale started out with 16 people. Soon, said Iandola, “we will be a company with 18 people. Most of them are engineers.”

Products in development
DeepScale is currently offering a reference kit to car OEMs and tier ones, so that they can try it out to refine their perception systems.

DeepScale is known for using efficient deep neural networks (DNNs) on small, low-cost, automotive-grade sensors and processors to improve the accuracy of perception systems that interpret and classify sensor data in real-time for automated vehicles. The goal for DeepScale is bringing driver-assistance and autonomous driving to mass-produced vehicles at all price-points.

DeepScale's approach: Deep Neural Network Sensor Fusion (Source: DeepScale)

DeepScale’s approach: Deep Neural Network Sensor Fusion (Source: DeepScale)

With input from tier ones and carmakers, Iandola expects his company’s reference kit to become mass production software ready for deployment into cars.

In parallel, DeepScale is hoping to develop its own testing methodology — similar to traditional functional safety methods but specifically designed for software-intensive vehicles. Iandola noted that traditional design/testing was developed for mechanically driven cars with a few electronics components. As new-generation vehicles bring in more software and AI, testing methodologies must change. “We hope to develop something by the end of this year and hope to share our methodology with customers and industry organizations.”

Partnership with Hella
Hella-Aglaiais a well-known tier one based in Germany, also sometimes regarded as a tier two supplying some key automotive components. By working with Hella, DeepScale’s CEO said, “We are committed to promoting an open platform,” as Visteon and DeepScale are already doing. Whether radars or vision chips are being supplied by Texas Instruments, NXP or Renesas, the real question tier ones face is, once the sensory data is ready to be fused, who writes that software? DeepScale hopes to be in the mix by providing efficient DNN for ADAS and autonomous driving software.

— Junko Yoshida, Chief International Correspondent, EE Times

Subscribe to Newsletter

Leave a comment