What’s actually limiting the speed of my deep neural network?

Deep Learning processor companies often highlight their products’ blazing-fast speeds in terms of metrics such as Tera Operations per Second (TOP/s) or Tera Multiply-Accumulate Instructions per Second (TMAC/s). What does this really mean, and are these numbers actually useful?

But first, what does this have to do with deep learning?

Let’s consider a convolution layer with 3x3x100 filters and 100 output channels.

• Let’s say that this layer has an input grid of size 50x50x100. …

Talk: Silicon Valley Deep Learning Group

Deep Learning is a broad field, and I view it as an interdisciplinary team sport. From my perspective, there are four main disciplines within Deep Learning:

• Data (including data annotations, data life-cycle, and data storage)
• Models (including the model structure, loss function, and quantization scheme)
• Infrastructure (every line of code that underpins the development, improvement, and deployment of deep neural networks)
• Applications (the motivating use-case for your deep neural nets)

In a high-output deep learning team, I’ve found that it makes sense to break down silos and to encourage people with these diverse skills…

We’re very happy to announce the closing of our Series A round of financing, with $15M raised to help us execute our plan to put deep learning into mass-produced cars and to make our roads safer. We’re excited to be partnering with our new lead investors, Point72 and next47. Other participants in our Series A round include Autotech Ventures and Trucks Venture Capital, two firms known for their automotive expertise. These funds are enabling us to grow our team, so if you’re interested in helping develop the world’s most efficient deep neural networks, contact us or browse our openings.

For more details on our funding round, see our CrunchBase page and our press release.

Until next time, feel free to follow us on Twitter at @DeepScale_.

Building on our previous posts covering the advent of general purpose computing in cars, and our approach to Deep Neural Networks, today we want to go further into the current status of computer vision in cars, and what we believe differentiates our approach.

The ADAS-Autonomous Spectrum

At one end of the spectrum in automotive vision are camera-only or camera+radar approaches, which are already widely used in ADAS (Advanced Driver Assistance Systems) features such as lane keep assist, adaptive cruise control, and automated emergency braking. Mobileye has emerged as a dominant player in these efforts. Mobileye went into business nearly 20…

Talk: ICML TinyML 2017

I’d like to give a big thanks to Manik Varma, Venkatesh Saligrama, and Prateek Jain for inviting us to speak at the ICML TinyML Workshop this year. They organized a great lineup of talks about efficient and resource-constrained Machine Learning, and it was a really fun event. They were also kind enough to record our talk:

The slides are available here: https://www.slideshare.net/ForrestIandola/small-deepneuralnetworks-their-advantages-and-theirdesign

I also want to thank Andrew Howard, Pete Warden, Song Han, and Ross Girshick for their helpful discussions on the content that led up to this talk.

Raw Data + Big Processor = New Possibilities

This is intended as background reading for those who are new to the Automotive industry.

In the past, most cars had few outward-facing sensors and modest computing power. These sensors were connected by a thin set of wires (such as CAN-BUS) that sent data slightly faster than a dial-up modem. In the last few years, that has changed dramatically.

The driver-assistance functions in vehicles, which were previously spread across dozens of microcontrollers, are now often implemented on a big processor called an ADAS Domain Controller. This change is not just happening at…

As CEO, a big part of my job is to scour the earth for the best engineers and entrepreneurs. After months of nonstop interviews, I am pleased to announce that several top-notch people are joining the DeepScale team.

Paden Tomasello — Software Engineer

Paden became passionate for HPC and ML while pursuing his undergraduate degree in EECS at UC Berkeley. While in school, he was awarded achievements on projects including parallel ray tracing optimization and computer vision using neural nets. His industry experience includes working at Cloudera, where he helped develop Impala’s JIT compiler, and Graphistry, where he scaled its core clustering algorithms using efficient…

At DeepScale, we are primarily focused on serving the automotive industry. We are a product-oriented company, and we will release more details of our product direction throughout 2017. Meanwhile, our product development efforts have required us to break new ground on core research problems in machine learning and deep learning. Where possible, we like to give back to the research community by releasing our new discoveries in the form of technical reports and open-source software. In this blog post, we summarize some of the things that we have released so far. …