New research could put a brain-like neural network in your smartphone

Stream:

News Bot

Your News Bitch
3,282
0
0
0
Console: Headset:
Mobile By Graham Templeton Mar. 20, 2014 9:58 am
Neural networks are popping up everywhere these days, and not just in the information-security complex that includes bodies like Google, Cisco, Facebook, and the NSA. That’s where neural networks have made the most headway so far, flooded as that industry is with powerful and versatile super-computers and staffs of coding geniuses to direct them. But now even non-moonshot corporations are getting in on the game.
Still, artificial neural networks, and the “deep learning” they can allow, have traditionally been inaccessible to the very individuals whose brains were the model for these networks in the first place. Deep learning has always been far too hardware- and even software-intensive to bring into the home.*Now, researchers at Purdue University claim that a combination of several innovations could bring deep learning to the common smartphone.
The basic advantage of a neural network is that it breaks complex problems down into stages. This is roughly similar to how (we think) the brain works: information is routed through a hierarchical organizational structure, moving through decision points and quickly sifting through extremely complex patterns by subjecting it to many quick, simple filters. As we grow and acquire knowledge, the physical and electrical connections in our brain can morph, making certain paths quicker and more efficient, or more accurate and detailed, or both. Similarly artificial neural networks are defined by their ability to integrate new pieces of information and refine its own decision-making process.
Researchers are working to enable smartphones and other mobile devices to understand and immediately identify objects in a camera’s field of view, overlaying lines of text that describe items in the environment. (Purdue University image/e-Lab)

The thinking machines that do this for Google are prohibitively expensive, but Purdue’s new technology can supposedly chew through images more than 120 times faster than conventional smartphone processors. Performance is supposed to be a full 15 times better than normal graphics processors, and the team predicts they could achieve another 10-fold increase with time.
On the back-end, this sort of functionality could change what a search engine does at a fundamental level, or be a quantum leap forward for Facebook’s ability to understand and sell your life. This research wants to put that functionality in your pocket, however, with a combination of custom hardware and software. The goal is on-board image analysis — that is, to have a smartphone run the analysis without having to farm the question out to some distant corporate server. This seems to undermine one of the great advantages of the neural network as a concept, however, since pooling the experience of all users would logically lead to a more detailed and efficient algorithm than any one user could compile.
This schematic diagram shows the basic idea of branching decision trees.

There are few details about the research as of this writing. We do not know, for instance, how its speed stacks up to a cloud-based solution, nor just how much “custom hardware” is needed to allow a “normal smartphone processor” to do these calculations. However, deep learning on a smartphone could allow us to make sense of unfamiliar objects — the developers provide the rather worrying example of identifying cancerous tissue inside a medical image. All dangers of computationally-assisted hypochondria aside, though, this could allow everything from on-board sign translation to facial recognition. You might ask your phone what sort of fruit you just found at at the market, or who this person is who’s just approached as though they know you.
These are definitely promising and potentially useful advances — but it remains to be seen whether the cloud will advance quickly and reliably enough to make these sorts of on-board solutions defunct before they even appear. Companies have an incentive to keep your data flowing through them, so even a useful technology might have trouble finding purchase in a developer- and manufacturer-controlled industry.



More...