Tesla Acquires DeepScale, a Startup Improving Deep Neural Nets for Self-Driving Cars

Home > News > Content

【Summary】Tesla, which becomes the latest automaker to buy a promising tech startup. The electric automaker announced it has acquired DeepScale, a computer vision startup that spun out of a research project at the University of California Berkeley.

Eric Walz    Oct 01, 2019 7:00 PM PT
Tesla Acquires DeepScale, a Startup Improving Deep Neural Nets for Self-Driving Cars

As the auto industry races to develop robust and affordable self-driving systems for mass production, automakers have been busy buying up small tech startups working on autonomous driving technology in order to jumpstart their own efforts. Now Tesla becomes the latest automaker to buy a promising tech startup. 

Tesla has acquired DeepScale, a California-based computer vision startup that spun out of a research project at the University of California Berkeley. The acquisition was first reported by CNBC

Two other people familiar with the deal confirmed to CNBC that Tesla had bought the company outright, but declined to disclose the terms of the deal.

Deepscale's technology may help Tesla improve its automated driving system Autopilot, which utilizes deep neural nets to make real-time driving decisions. DeepScale is working on artificial intelligence perception software for driver-assistance and autonomous driving, with a focus on implementing efficient deep neural networks on low-cost, automotive-grade processors. 

DeepScale co-founder and CEO Forrest landola completed his doctorate at UC Berkeley working on deep neural networks and computer vision systems. While working with his faculty advisor and professor Kurt Keutzer, landola's advances in scalable training and efficient implementations of deep neural networks (DNNs) led to the founding of DeepScale. Keutzer became co-founder of the new company.

The week, landola announced that he joined Tesla as a senior staff machine learning scientist and is now working on Tesla's Autopilot team.

DeepScale is Improving Autonomous Vehicle Perception Using Deep Neural Nets

DeepScale's unique solution for a more efficient perception systems includes using DNNs on small, low-cost, automotive-grade sensors and processors to increase their accuracy. The DNNs interpret and classify sensor data in real-time so an autonomous vehicle can safely navigate.

DeepScale's deep neural nets (DNNs) analyze data from various sensors to help autonomous vehicles perceive the world around them by identifying objects such as pedestrians and other vehicles, which is a common application of DNNs used by self-driving car developers. 

The perception technology developed by Deepscale collects raw sensor data instead of object data, and then use a low-cost embedded processor to accelerate sensor fusion in order to make sense of it.

Much of the current research on deep neural networks (DNNs) has focused on achieving a high-level of accuracy. However, a significant amount of work and research has been conducted on distributed training of neural networks at scale. 

Deepscale discovered that to achieve the desired accuracy level, it is possible to use multiple DNN architectures which have lower computing demands, but still achieve the same high accuracy needed for autonomous driving, according to the company.

Using smaller DNN architectures offers two main advantages. First, smaller DNNs require less communication across servers during distributed training, which reduces latency. Second, for autonomous driving applications, smaller DNNs require less bandwidth to export a new model from the cloud to an autonomous car using over-the-air updates. 

While most modern neural network architectures trained on large data sets can obtain impressive performance, these neural network models are computationally demanding. In addition, Network training can take much too long using a single machine.

To provide all of these advantages, DeepScale developed a small DNN architecture called SqueezeNet. SqueezeNet achieves greater accuracy identifying objects from images using 50x fewer parameters. In addition, DeepScale was able to compress SqueezeNet to less than 0.5MB of space, making it ideal for use in the auto industry.

Before being acquired by Tesla, DeepScale was working on ways to get SqueezeNet to run on inexpensive hardware and uses less power, the same goal Tesla has with the latest version of Autopilot.

Tesla built its own full self-driving (FSD) hardware after after finding that current processors from companies like NVIDIA did not meets Tesla's computational needs. Tesla's acquisition of DeepScale could help Tesla to improve its Autopilot perception capabilities even further and at a lower cost.

The acquisition of DeepScale is Tesla's second major purchase this year. In February, Tesla bought Maxwell Technologies for $218 million, a San Diego-based developer and manufacturer of battery solutions, including ultracapacitors and electric vehicle battery cells. 

Ultracapacitors show promise for use in electric vehicles. They store energy in an electric field, rather than in a chemical reaction like traditional batteries, allowing them to charge and discharge almost instantly, delivering rapid surges of energy. Tesla CEO Elon Musk is said to be a fan of the technology and its potential for use in electric vehicles.

Prev                  Next
Writer's other posts
    Related Content