PORTLAND, Ore. — Biologically inspired neural processing units (NPUs) were recently described by Qualcomm Inc. in San Diego at the MIT Technology Review's EmTech conference. Qualcomm chief technology officer (CTO) Matt Grob described a new generation of NPUs and design tools that Qualcomm hopes to make available to developers next year.
Purdue University researchers also use neural networks to create an image-processing application that can categorize objects from a moving car in real-time.
(Source: Eugenio Culurciello, Purdue University/Qualcomm at MIT's EmTech)
At the conference, Grob showed videos of what it calls its Zeroth Robot prototype -- named after Isaac Asimov's Zeroth Law of Robotics (that no robots shall harm humanity). These robots were not powered by a conventional computer but instead by biologically inspired NPUs modeled on the human brain and created in cooperation with Brain Corp, which receives funding from Qualcomm Ventures and operates its labs inside Qualcomm's facility.
According to Grob, even though these early prototypes are general-purpose image processors that learn their application rather than depend on complicated hand-written algorithms, they are already offering comparable performance to the best custom-designed image processing algorithms for conventional computers today. Qualcomm has also cited professor Eugenio Culurciello, who is using Purdue University's own neural network development tools, to perform real-time image recognition of objects from moving cars (see figure).
Grob promised that after proving its neural processing units in robotics applications, Qualcomm envisions using these chips in its core business: mobile handsets. He described applications for cellphones that offer more natural interfaces, where the user trains the phone rather than being forced to learn complicated menu commands. Qualcomm also aims to incorporate sophisticated, neural-based search capabilities through big-data that are very efficient compared to the power-hungry remote servers used today.
"Mobile is a very challenging design environment, we are under constraints for power, performance, size," said Grob at EmTech. "And it turns out a brain is an incredibly high-performance system with these same features -- very power efficient -- with incredible density of performance when you consider what it's doing."
The neural difference
Today all commercial processors are based on the von Neumann architecture, which separates processing from memory, where they compute and store results. Modern processors offer some small modicum of parallelism using multiple cores, but they are still based on the same antiquated principles -- the so-called Harvard architecture defined at the dawn of the age of computers.
"A brain is nothing like that, so we are looking to biology to inspire us for a new generation of processors," said Grob at EmTech. "The brain possesses superior capabilities for image recognition as well, so we are trying to understand why that is and bring that to bear."
Instead of performing a million processing steps with less than 10 parallel execution units, as is done by today's multi-core processors, the brain does the opposite -- it solves the same problems by performing less than 10 processing steps but with a million parallel execution units, according to Grob. The brain is also very power efficient, he explained, consuming only about 20 watts at a cost of under a quarter of a cent per hour, whereas simulating the brain on a conventional von Neumann computer would take up to 50 times more power.
To replicate what the brain is doing, Qualcomm, in cooperation with its Brain Corp., has developed a spiking model of the visual system. Since 2009, Brain Corp. has been perfecting its models of the spiking behavior of neurons, creating models of their transfer functions that replicate its biological behaviors in computationally efficient ways. Their strategy was to create a suite of neural network application development tools that model the behavior of neural networks, that the company will soon be providing to application developers.
"Our suite of tools goes all the way from design synthesis and simulation to realization in hardware," said Grob. Besides using its tools internally to create smarter cellphone chips that can learn the habits and preferences of their users without having to be explicitly configured, Grob envisions all types of other products making use of its forthcoming neural network microchips.
For instance, toy radio-controlled (RC) vehicles could be made available that have already learned from world famous race car drivers how to avoid obstacles. Qualcomm also envisions alternatives to app stores, which he called "experience stores," allowing users to download expertise into their consumer products.