Compute sticks aren’t new, but a compute stick that brings neural networks to other devices, is. Movidius, a firm specialising in embedded machine vision, has today announced the Fathom Neural Compute Stick. Movidius claims that it’s the world’s first embedded neural network accelerator. Fathom is a device that looks very similar to the Intel Compute Stick, and you may even be forgiven for thinking it does the same. Actually, it’s meant to do so, but only for devices that need to add machine learning, or neural networks.
To simplify, the Fathom is meant to make drones, cameras etc. smarter. The Stick has a Myriad 2 chip, from Movidius, inside. Plugging it into a drone, or a Go Pro, would allow that device to take advantage of the Myriad 2’s capabilities. What does the Myriad 2 do? Well, this is the chip running on DJI’s Phantom 4 drone. Movidius’ chips have also been an important part of Google’s Project Tango, which brings smart augmented reality to devices. Getting the point now?
Movidius has also developed the Fathom Deep Learning Software Framework, for developers to build using the Fathom. The Myriad 2 inside, is a low power chip, designed specifically for smart computer vision in devices like drones, smartphones and more, where power consumption is a big concern.
Moreover, by using a Myriad 2, a developer can essentially add intelligent systems to any qualifying device. This includes things like drones, GoPro, and even tablets. Instead of building their own learning systems, and then using the cloud to get that onto a device (like a drone), the Fathom can allow such tasks to be performed in real time. This marks a considerable step in bringing machine learning to the mainstream.
The Fathom’s use cases can range anywhere from smart obstacle sensing drones, to better speech recognition on devices. Of course, the device is meant for developers at the moment, and Movidius is offering 1000 units for free to “qualifying customers”, at the moment. The company plans to start selling the Fathom around the end of the year, at a sub-$100 (approx. Rs. 7000) price point.