jeudi 24 octobre 2013

Qualcomm to Build Neuro-Inspired Chips

A lire sur:  http://www.technologyreview.com/news/520211/qualcomm-to-build-neuro-inspired-chips/

World’s largest smartphone chipmaker offers to custom-build very efficient neuro-inspired chips for phones, robots, and vision systems.
Chips and software that draw on the brain’s ability to do massive parallel processing could make for far smarter, more power-efficient technologies.

The world’s largest smartphone chipmaker, Qualcomm, says it is ready to start helping partners manufacture a radically different kind of a chip—one that mimics the neural structures and processing methods found in the brain.
The approach is emerging as a way to enable machines to perform complex tasks while consuming far less power. IBM has been prototyping similar chips (see “IBM Scientists Show Blueprints for Brainlike Computing),” and the area is the focus of intense research around the world (see “Building a Brain on a Silicon Chip” and “Intel Reveals Neuromorphic Chip Design”).
Speaking in a sponsored talk at MIT Technology Review’s EmTech conference today, Qualcomm CTO Matt Grob said that by next year his company would take on partners to design and manufacture such chips for applications ranging from artificial vision sensors to robot controllers and even brain implants. The technology might also lead to smartphones that can sense and process information far more efficiently.
Today’s computer systems are built with separate units for storing information and processing it sequentially, a design known as the Von Neumann architecture. By contrast, brainlike architectures process information in a distributed, parallel way, modeled after how the neurons and synapses work in a brain.
Qualcomm has already developed new software tools that simulate activity in the brain. These networks, which model the way individual neurons convey information through precisely timed spikes, allow developers to write and compile biologically inspired programs. Qualcomm is using this approach to build a class of processors called neural processing units (NPUs). It envisions NPUs that are massively parallel, reprogrammable, and capable of cognitive tasks like classification and prediction. “What we’re talking about is scale, making it into a platform,” said Grob during his talk. “We want to make it easier for researchers to make a part of the brain.”
For several years Qualcomm and Brain Corp, a separate company it has invested in, have been working on hardware and algorithms that attempt to mimic the processes of the human brain. The company calls the overall program Zeroth, borrowing from the science fiction author Isaac Asimov’s “Zeroth Law of Robotics” (which specifies that robots must not harm humanity).
The company has built prototypes of a neuro-inspired chip for a rolling robot in Qualcomm’s labs in San Diego. Simply telling the robot when it has arrived in the right spot allows it to figure out how to get there later without any complex set of commands, Grob said (see video).
“This ‘neuromorphic’ hardware is biologically inspired—a completely different architecture—and can solve a very different class of problems that conventional architecture is not good at,” Grob said in an interview after his talk. “It really uses physical structures derived from real neurons—parallel and distributed.”
He added, “What is new now is the ability to drop down large numbers of these structures on silicon. The tools we can create are very sophisticated. The promise of this is a kind of machine that can learn, and be programmed without software—be programmed the way you teach your kid.”
Grob said that by next year the company will be able to partner with researchers and startups, offering them a platform to realize designs in hardware. “We are bringing the ability to develop it and scale it, too. We bring scale to the picture,” he said. “They can create neural architectures very quickly with our tools, and simulate them, and take it all the way to realization.”
Aviva Rutkin contributed reporting.

Aucun commentaire:

Enregistrer un commentaire