Word on the street is that Tesla has now equipped its ever-growing tech-arsenal in the form of DeepScale. This UC Berkeley computer vision startup has proved their mettle last year with their work on SqueezeNet and Carver 21.
The news of this acquisition was followed by DeepScale CEO’s Forrest Iandola post on social media announcing his association with Tesla as a senior staff machine learning scientist.
— Forrest Iandola (@fiandola) October 1, 2019
Tesla has been building a unique portfolio of highly regarded engineering companies. Earlier they have bought companies like Grohmann Engineering, a German company and a supercapacitors manufacturer, Maxwell Technologies.
Tesla enjoys the advantage of having started their journey very early. They possess the world’s largest customer base for semi-autonomous vehicle. Whenever a Tesla driver takes an action, be it steering left or right or pressing the pedal, what they are doing is annotating the data and generating more refined data related to the driver’s behaviour.
With time, more data will be generated and every possible driver reaction will be captured. This kind of data requires a computational efficiency of the highest order.
Now with the acquisition of DeepScale, Tesla looks to fill in the computational voids associated with computer vision tasks.
What DeepScale Has To Offer
DeepScale came into limelight with the publishing of its work, SqueezeNet, as the name suggests is an attempt to squeeze every ounce of computational energy from deep neural networks. The significance of this work is immense considering the kind of data the on-board cameras of a self-driving car generates.
Convolutional neural networks are the workhorses of modern-day computer vision models.
At the fundamental level, these operations are nothing but a bunch of multiplications and additions (dot product) happening on a large scale. A human-annotated dataset may not capture all the complexity of the real-world scenario in real-time. So, the neural nets have to be automated for training as the data gets streamed through onboard cameras.
This is where DeepScale’s SqueezeNet comes into the picture. With its accuracy matching its state-of-the-art counterparts, it enabled efficient training of neural networks, which are desirably lighter.
SqueezeNet was able to match AlexNet’s accuracy on the ImageNet benchmark with fewer parameters, and one of its deeper variants is able to achieve VGG-19 accuracy with only 4.4 Million parameters, smaller than VGG-19).
It also achieved better top-5 classification accuracy with fewer parameters as compared to MobileNet.
SqueezeNet’s wide range of accuracy allows the user to make speed-accuracy tradeoffs, depending on the available resources on the target hardware.
After their exemplary work on efficient neural networks, the handful of bright machine learning team at DeepScale came up with Carver21.
Carver21, was aimed at enabling autonomous carmakers to build AI building blocks for their cars. With Carver21, DeepScale offered a portfolio of software modules that gave the much-needed flexibility for building scalable advanced driver-assistance systems (ADAS).
DeepScale’s full-stack deep learning methodology enables cohesive integration of AI software with various processors and sensors for customizable automated driving features.
The availability of a full-stack deep learning module takes care of every aspect from training, development, deployment, and even data collection/curation to produce proprietary state-of-the-art AI solutions.
Tesla Moves Ahead In The Race
A couple of years ago, Tesla announced that it would be shipping all of its cars with the necessary hardware to support the future advancements in autonomous tech.
And, today, as promised, Tesla not only managed to live up to their ambitious goal but also designed the world’s most advanced chip.
The real world is as weird as it can get and more so for driving amongst a bunch of other drivers with varying moods and motivations.
Real-time decision making is critical for any driverless vehicle. The ability to deploy more deep learning functions with higher accuracies in centralized high-compute environments is what any self-driving car manufacturer desires. And, deploying multiple DNNs in this tiny compute footprint is what made DeepScale’s solutions a right fit for Tesla’s aspirations.
The point is not to necessarily consume as little of a powerful computer as possible, but to maintain automotive-level accuracies while using fewer resources.
Tesla’s cutting edge technology combined with their portfolio of diligent acquisitions gives them that extra edge to keep the dream of full autonomy within touching distance.