Zeroth ‘brain chip’ technology comes to Snapdragon handsets

Qualcomm may be under pressure in its core markets (see separate item), but nobody discounts its ability to shift the goalposts with the power of its engineering capabilities. It has repeatedly, when under pressure, managed to define a new category of functionality or performance, creating a new headstart for itself in chip sales and IPR licensing. A recent example is Zeroth, its ‘brain chip’, with which it hopes to stake a claim in emerging cognitive computing platforms.

Zeroth was described last year and was on show at Mobile World Congress, in experimental form at least. It is not actually a chip but a software platform designed for future Snapdragon processors.

It is trying to do the opposite of all the massively powerful processors which are being developed to run cloud-based deep learning engines, and instead, move that artificial intelligence to the seat of Qualcomm’s power, the smartphone. Raj Talluri, SVP of product management, said in an interview in Barcelona: “We wanted to see if we could build deep learning neural networks on devices you carry with you instead of in the cloud.” Of course, the handset will not have the processing power to solve complex challenges like chess games, but it could use AI techniques for more everyday tasks such as handwriting recognition, or helping the camera distinguish between different types of objects, and optimize the shot accordingly. So far, Zeroth can recognize about 30 classes of objects using machine learning, and can support applications such as recognising the faces of friends in a photo and automatically linking them to Facebook.

Zeroth is described as a neuromorphic chip platform, which claims to mimic the way the brain works. While commercial application in 2015-16 will focus on enhanced capabilities for Snapdragon, eventually Qualcomm foresees the creation of a widely applicable ‘neural processing unit’ (NPU) to perform the core functions to support activities such as deep learning, in the same way that a GPU supports key underlying functions to drive graphics. It feels like a long term bet ‘ and an area where Intel is also active, but the pay-off could be huge if neural networking, after so many false dawns, becomes mainstream in the internet of things, and if Qualcomm establishes the leading platform to deliver it.

The first public demonstration of Zeroth, reported by MIT Technology Review, featured a small robot called Pioneer, which rolled up to a Captain America action figure, picked it up with a scoop and pushed it towards three pillars. The robot ‘saw’ a Qualcomm engineer pointing to one of the pillars, and set the action figure down in that place. It then took a second figure and puts it in the same place, having ‘learned’ the right location. It was also able to distinguish between the dolls and other objects, which it ignored. The difference between this and object recognition technologies is that the robot also takes action based on what they see.

“If you and your device can perceive the environment in the same way, your device will be better able to understand your intentions and anticipate your needs,” Samir Kumar, a business development director at Qualcomm’s lab, told MIT Technology Review, while CTO Matthew Grob commented: “We’re blurring the boundary between silicon and biological systems.”