Assume you want a server farm to hold out AI-powered duties like facial and object recognition? Assume once more.
The $35 Raspberry Pi has lengthy been able to working picture recognition software program, with hobbyists utilizing the board to pick the whole lot from faces in videos to obstacles in a robot’s path.
Now the Intel Movidius Neural Compute Stick (NCS) guarantees to spice up the speed at which the Pi can perform vision-related duties like facial and object recognition.
The $79 stick plugs into the Pi through USB, and accelerates imaginative and prescient recognition duties utilizing the 12 specialised cores in its Myriad 2 Imaginative and prescient Processing Unit (VPU). The low-power processor is able to 100 gigaflops and consumes a single watt, though the facility draw for the stick might often rise to 2.5W.
Movidius has now released a video, see above, and text guide demonstrating check out object recognition on the Raspberry Pi three and the NCS, displaying off the system recognizing sun shades and a pc mouse as a digicam pans round a room. Getting this demo working requires downloading a couple of software program libraries, proudly owning a Pi digicam and copying some information from a PC working Ubuntu 16.04.
An early model of what would turn out to be the NCS was introduced by Movidius final April, then a prototype machine referred to as the Fathom, however, after Intel purchased Movidius in September 2016, it was by no means totally launched.
On the time, Dr Yann LeCun, Fb’s director of AI analysis and founding father of convolutional neural networks, described the Fathom as a major step ahead.
“As a tinkerer and builder of assorted robots and flying contraptions, I have been dreaming of getting my arms on one thing just like the Fathom Neural Compute Stick for a very long time. With Fathom, each robotic, large and small, can now have state-of-the-art imaginative and prescient capabilities,” he said. The specs of the 2 sticks are broadly the identical, with the primary distinction being the NCS has 4GB of reminiscence, 4 instances that of its predecessor, to assist denser neural networks.
Intel says the Movidius NCS will assist decrease the limitations for individuals who wish to get began with deep studying utility improvement and supply a easy approach so as to add visible recognition techniques to prototype gadgets, reminiscent of drones, surveillance cameras and robots.
Since the entire knowledge is dealt with by a regionally saved neural community, the NCS would not require an web connection. It appears to be suited to make use of instances the place the latency of speaking with a server can be too nice, the place privateness is a priority, or the place a high-performance processor can be too energy hungry.
The NCS would not speed up the computationally intensive course of of coaching a neural community to hold out imaginative and prescient recognition, which can sometimes nonetheless require a extra highly effective laptop than the Pi.
As a substitute the NCS boosts the pace at which pre-trained networks make inferences about knowledge, for instance whether or not there’s a cat in a picture. Tough estimates of performance online say the stick’s VPU can do 10 inferences per second utilizing a GoogLeNet convolutional neural community, in comparison with about 2 inferences per second utilizing Google’s Inception convolutional neural community structure on an unaided Raspberry Pi.
Movidius says customers can chain a number of sticks collectively, which can ship a linear increase to efficiency for every stick added.
At current, the NCS solely helps machine-learning fashions constructed utilizing the Caffe framework, though Movidius has hinted that support for neural networks built using Google’s TensorFlow library could also be on the playing cards.