DeepScale
| File:Deepscale logo.png | |
| Founded | 2015 |
|---|---|
| Founder | Template:Plain list |
| Defunct | October 1, 2019 |
| Fate | Acquired by Tesla, Inc. |
| Headquarters | Mountain View, California, U.S |
Key people | Forrest N. Iandola (CEO) |
| Website | deepscale |
DeepScale, Inc. was a U.S.-based technology firm located in Mountain View, California that specialized in perceptual systems for automated vehicle technologies. The company was acquired by Tesla, Inc. on October 1, 2019.
History
DeepScale was established in 2015 by Forrest Iandola and Kurt Keutzer. In 2018, the company completed its Series A financing round. That same year, it entered into strategic collaborations with automotive suppliers such as Visteon and Hella Aglaia Mobile Vision GmbH. The 2019 acquisition by Tesla was consistent with Tesla’s commitment to enhancing autonomous driving technologies.
Technology
Prior to creating DeepScale, Iandola and Keutzer worked together at University of California, Berkeley focusing on optimizing the efficiency of deep neural networks (DNNs). Their joint effort in 2016 resulted in the development of SqueezeNet, a compact and power-efficient DNN designed for computer vision. This breakthrough allowed deep learning algorithms to operate effectively on devices with limited power resources, including smartphones and automotive-grade processors. By 2018, DeepScale had advanced beyond SqueezeNet by producing faster and more precise DNNs aimed at commercial automotive perception solutions.
Neural architecture search
Neural architecture search (NAS) has recently outperformed manual design methods in producing DNNs with superior accuracy and speed. In 2019, DeepScale unveiled SqueezeNAS, a NAS-based supernetwork technique that generated rapid and precise DNNs for image semantic segmentation tasks. The SqueezeNAS models surpassed the performance tradeoffs seen in Google's MobileNetV3 architectures. Importantly, while Google’s development of MobileNetV3 required thousands of GPU-days, DeepScale’s approach completed comparable designs within just tens of GPU-days.
Product
DeepScale created perceptual system software powered by deep neural networks to enable vehicles to understand their environment. This software operates on an open platform compatible with a variety of sensors and computing units, ranging from NVIDIA GPUs to ARM processors designed specifically for automotive applications.
In January 2019, DeepScale introduced "Carver," an automotive perception solution employing three simultaneous neural networks responsible for object detection, lane recognition, and drivable space identification. Carver functions in real-time with a processing throughput of 0.6 tera-ops per second. For reference, each chip within Tesla’s Full Self-Driving computer delivers 36 tera-ops per second, meaning Carver requires approximately 2% of one Tesla chip’s computational capacity.
Acquisition by Tesla
The acquisition of DeepScale by Tesla was announced on October 1, 2019, as reported by CNBC. According to Fortune, DeepScale’s technologies were slated for integration into Tesla’s Autopilot platform. Furthermore, CNET emphasized that DeepScale’s approach to autonomy complements CEO Elon Musk’s vision of relying on cameras, radar, and ultrasonic sensors instead of LiDAR to build a resilient self-driving system.