Slashdot Mirror
Google Puts Souped-Up Neural Networks To Work
holy_calamity writes "A machine learning breakthrough from Google researchers that grabbed headlines this summer is now being put to work improving the company's products. The company revealed in June that it had built neural networks that run on 16,000 processors simultaneously, enough power that they could learn to recognize cats just by watching YouTube. Those neural nets have now made Google's speech recognition for U.S. English 25 percent better, and are set to be used in other products, such as image search."
How long until Google starts using this for face recognition? It already has all the images of the internet indexed and is trying to make their own social network Google+ with profile pics and hidden party photos. How long until Google starts using this to recognize faces "to improve targeted ads"?
And just think about all the possibilities with facial recognition. Endless amounts of goodness right? Database of faces
Actually Siri sucks and I hate it
The most useless feature I've tried to use
How long before this neural network will be able to steal banking credentials and funnel money to finance its own army?
What is the proposed name of this, ehm, highly innovative product?
Google God (Beta)
In AI circles, a popular saying is that Neural Networks are always the second best way to solve a problem. Its what you use when you don't want to (or don't know how to) implement a more specific approach.
"His name was James Damore."
I would think the first thing they should do would be put neural networks to use learning how to build better neural networks, then use the improved version for the same process.
If anybody is curious what the breakthrough is, it's just having networks learn by trying to recreate their input data (as opposed to trying to figure out some 'right answer'). Pretty obvious really.
In today's news, google announced that a new algorithm has achieved a 90% success rate in identifying video's containing cats on youtube. The algo shouts "cats" every time a video is started, and since 90% of youtube video's contain cats, the algorithm has obtained a success rate of 9 in 10.
Google 'Iris'.
"Aineko", or "Beautiful cat".
Google's main product is supplanting ECHELON for the NSA, so well done for making that more productive. Thanks from (the rest of) the developed world..
"My CPU is a neuro net processor, a learning computer."
The Internet King? I wonder if he could provide faster nudity.
Google it yourself sjeeee.
enough power that they could learn to recognize cats
How many more nodes can they add before it wants to know what they taste like?
Everybody's beating around the bush on this so I'm going to say it right out in the open, this sounds like skynet...
The image of the ideal face stimulus in the article
http://www.technologyreview.com/files/92225/google.machine.learning2.jpg
contradicts the head tilt research in this paper:
http://www3.canisius.edu/~noonan/research/researchreports/human_head_tilt.htm
Or maybe we tilt the camera further to the left than our heads to the right...
Google 9000. Take that IBM. And it will not have to go to Jupiter to reveal itself.
Artifical Neural Networks are extremely expensive and we don't really know how they work.
Compared to that, natural neural networks are moderately expensive (to the large corpos) and after some training quite efficient. We sometimes think we know how they work. At least the psycho-science says so.
I always say "how stupid can it be" when they fantasize about replacing a well-trained 100-billion neuron biological supercomputer by a brittle satcom link and a 700-neurons "AI neural network" in those imagined drone fighters. That 100 billion supercomputer can do much more complex decisions than the 700 neuron thingy (that is what we can currently pack into a drone, realistically), for very obvious reasons. Just make sure the 100 billion thing gets good training, O2, food and from time to time a complementary brain to fuck. Stick some colorful patches on the physical hull of that supercomputer and it will do the most insane things, including killing itself for the objective. They call it the "Kamikaze supercomputer". And it does not need a fecking SATCOM line in case you are up against civilized forces and not just AK47-wielding Neanderthals who can't do the most basic SIGINT and EWAR.
Any electrical engineer worth their salt can jam these drone SATCOM links with less than 500$ of parts and less than 5000 dollars worth of T&M/monitoring equipment. Same with recce sats of all kinds. I've seen it. Lots of nasty stars instead of crisp 0.5-meter resolution images.
But who said the military is a rational affair ? It's a scam for the benefit of uniformed politicos and their buddies in the MIC. Who are the same by means of revolving door.
As we don't really know what goes on in a reasonably complex artificial NN, they are very dangerous from an engineering perspective. They can surprise you any time. It is more Voodoo than solid engineering. As Google is largely an adbroker and lots of voodoo, I could not care less. Just make sure they don't control any live
cars with NNs.
That is what a guy from a major automotive company's research org told me 15 years ago. They had autonomous cars driving all over Germany back then, but they would not use NNs, as they are opaque and cannot be trusted to properly work under all conceivable conditions. They invented the car (not just the autonomously driving one) and they do know some stuff.
Very much like these plasma spheres which make very nice looking gas discharge glows, depending on where you put the hand. Nice to look at, not fit for any critical purpose, as there are no proper theories about it.
Why do you use Google Mail, why do you use it w/o an anonymizer, why don't you purge your cookie list every day ?
Because you are a lazy fuck who would trade your mother for convenience.
Google is the NSA Front Office for all the cretinous people who like to be assfucked and controlled 100% of time. Those who eat all the TERRORIZM propaganda. Those who hand-wring about Nazi-time warfare, but who happily eat the whole "need some war quickly" propaganda menu of the Virginia Mafia and the Jerusalem Crooks.
And thus Skynet was born..
90% of 90% would be 82% accuracy, plus 9% false negatives, plus 9% false negatives:
---------------x
xxxxxxxxx-
xxxxxxxxx-
xxxxxxxxx-
xxxxxxxxx-
xxxxxxxxx-
xxxxxxxxx-
xxxxxxxxx-
xxxxxxxxx-
xxxxxxxxx-
why the we use the phrase 'souped-up'? I could understand 'suped up' as in 'super' but, to misquote Tina Turner, what's soup got to do with it?
I'm surprised because a few years there was a Google Tech Talk by Geoff (sp?) Hinton. He demonstrated a few techniques and showed excellent results in computer vision and document categorisation.
Although maybe Google have been using Neural Networks for years, and are only now hearing about this slightly silly use.
Buuut, you seem to have misread. 90% of 100%. 100% of the videos are called "cat" videos. 90% are correct. 10% are false positives.
I know that if I were a mad scientist working Google, the first thing I'd do would also be to build an artificial sentience and show it mankind's collection of cat videos. I mean who wouldn't?
When will this neural network be put to the task of building more effective neural networks?
The 80s are over, so why use tech from then? I could think of a dozen techniques (SVMs, baysian nets, ELMs...) that routinely beat the crap out of neural networks in terms of both accuracy and practicallity. In fact the only real advantage I can see for NNs is the cool sounding name. So why?
(and yes, I a a researcher in machine learning)
Soylent Google I believe.
BlameBillCosby.com
This explains who was watching all of those cat and kitten videos on Youtube.
~ Best man at your service.
I've written signal processors that can detect different flame noises, and I've written video software that detects smoke movements.
Using emotive language and writing like this just blurs the hell out of the problem space - you have to anthropomorphize a computer before you can imagine these possible solutions?
We've tracked footsteps in 3d in a room from several people using a few mics - there are lots of applications, you don't have to pretend you're using some "meta-programmed neural net" that can understand what you want.
Is Google run by a neural network?