Put AI to work
June 26-27, 2017: Training
June 27-29, 2017: Tutorials & Conference
New York, NY

Software and hardware breakthroughs for deep neural networks at the edge

Michael B. Henry (Mythic)
11:55am12:35pm Thursday, June 29, 2017
Implementing AI
Location: Beekman Level: Beginner
Secondary topics:  Deep Learning, Hardware, IoT and its applications

Prerequisite Knowledge

  • A general understanding of AI and where it has currently achieved market penetration

What you'll learn

  • Understand how AI will make its way into everyday products within the next 3–5 years


How do you turn any product into a smart product? Have it engage with the user, understand the environment, be instantly responsive, and have exceptional reliability.

Breakthroughs in deep learning and new analog-domain computation methods to deploy trained neural networks will deliver exciting new capabilities to consumer and industrial hardware products. Michael B. Henry explains why the combination of human-like levels of recognition and massive computation capabilities in a tiny package will enable products with true awareness and understanding of the user and environment.

First key: Local AI—Cloud processing is unreliable, has poor latency, and imparts onerous data transmission requirements. Intelligence should be at the edge: in the device and as close to the user and environment as possible.

Second key: Pick the best algorithms—Deep neural networks (DNNs) are the best at interpreting data pulled from the real world (audio, video, sensors), and breakthroughs promise to broadly expand their scope. Generative/adversarial networks now give developers a tool to train networks using loss functions inspired by human perception. WaveNets allow us to separate noise and recreate rich data like voice with a stunningly low number of parameters. It’s not just about voice assistants and self-driving cars. Any device that requires user interaction or perception of the environment will reap huge benefits from the ongoing advances in deep learning.

Third key: Big improvements on every metric that matters (low cost, low power, low latency, high throughput)—To run ensembles of 1M–100M parameter networks that are instantly available yet disappear when not needed, a new method of running already-trained DNNs is required. Computation is done in the analog domain and inside of flash memory arrays. Each transistor in the array stores a DNN parameter and performs the computation. These techniques offer a 100x improvement over conventional digital computation, while the nonvolatile nature of the compute + storage lets you turn on intelligence as needed and save power when not.

Photo of Michael B. Henry

Michael B. Henry


Mike Henry is cofounder and CEO at Mythic, an AI hardware startup based in Redwood City, CA, and Austin, TX. Under his leadership, Mythic has raised $56M in investment from top-tier VCs and developed novel chip technology that is up to 100 times better than incumbents. Mike holds both a BS and a PhD in computer engineering from Virginia Tech.