Miscellaneous

Movidius’ Fathom – A Deep-learning Accelerator On A USB Stick

It’s not often a unique product shows up in our email and news feeds, but we became curious over something a little different today: a USB compute module for neural networks and machine learning. Movidius’ new Fathom Neural Compute Stick is an ultra-low power processor featuring a Myriad 2 VPU (vision processing unit), providing up to 150 GFLOPS of compute for about 1 Watt of power.

Neural networks and machine learning have fascinated people for a long time; teaching machines and computers to think rather than simply ‘do’. The problem is that it’s extremely difficult; not just from a hardware perspective (people aren’t binary), but from an academic and psychological perspective too; the how we think.

Instead of writing code that goes through images or live camera feeds, doing edge detection, writing more algorithms to define shapes and generally teach the computer what it needs to do (explicitly), neural networks are a way to teach a computer to teach itself – often creating systems that are faster and more accurate than the hand-coded method.

Over the last few years, neural networks and machine learning have hit the media’s attention a few times, usually when it goes wrong; but it’s been in the background of major businesses and software features for quite a while, almost completely transparent to the user. Face detection on Facebook, reverse image search, music finding by humming, even the natural language element of voice detection with Siri and Cortana.

The underlining element with all these features is that they are both powered by neural networks (be it software or hardware), and pretty much all of them require huge amounts of resources using cloud services to function. Siri and Cortana might seem instant, but they require an Internet connection for a reason, at least with their more advanced features.

Fathom

Sometimes though, that cloud compute connection isn’t always possible, and trying to pack in enough performance into a small package like a mobile or drone, can be challenging from a couple of fronts; power and heat. Movidius has released its Myriad 2 VPU and Fathom to help pioneer a new wave of portable machine learning. It goes without saying that companies like Facebook and Google are both interested in applications of this technology. The Myriad 2 chip is already available in FLIR systems for thermal imaging and the Phantom4 drone. Google will also be making use of the chip in upcoming deep-learning devices.

https://www.youtube.com/watch?v=XfCdArXDDKM

Deep-learning can be run on standard CPUs, but it’s a rather intense task, and specialized in nature. There’s a reason why NVIDIA has dedicated hardware projects for it with its Tegra TX1 system. By providing a simple USB stick with support for both major deep learning frameworks, TensorFlow and Caffe, Fathom allows for visual processing to be done on mobile systems directly, invaluable for drones and robots. This also frees up the main CPU for other tasks. It also allows for mobile development out in the field, without having to lug around a much larger and more expensive system (say, a Tesla P100).

Drones are one of the easiest ways to show the benefits of visual processing – terrain geometry calculation, collision avoidance, mapping, distance calculations, or the slightly more creepy, face detection and following. Movidius is also working on ways to integrate the Fathom with the GoPro camera. Instead of running pre-compiled algorithms on the device, these could be done on the Fathom, while letting it continue to teach itself as it flies. In fact, it could teach itself to fly better, as well, with real-time compensation for sudden updrafts.

The new Fathom Neural Compute Stick will be made available from Movidius for around $100 or less.

 

Source: TechGage

Related posts

Leave a Comment

* By using this form you agree with the storage and handling of your data by this website.

This site uses Akismet to reduce spam. Learn how your comment data is processed.

This website uses cookies to improve your experience. We'll assume you're ok with this, but you can opt-out if you wish. Accept Read More