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."
What is the proposed name of this, ehm, highly innovative product?
Please move back into the reality distortion field.
If Pandora's box is destined to be opened, *I* want to be the one to open it.
August 4, 1997. It's running very late.
"I bless every day that I continue to live, for every day is pure profit."
How long until Google starts using this for face recognition?
That totally does not bother me. These methods are easily defeated by Burqa technology, invented by Muslims ages ago.
If Pandora's box is destined to be opened, *I* want to be the one to open it.
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."
You're joking right. Google and Apple both already have face recognition software for years. The government has been using face recognition software for years also. Using face recognition to give targeted adds will happen one day, but the infrastructure to do that at your local mall isn't there yet. NN can help improve face recognition software, though it's not really necessary. Plus it's rather easy to fool face recognition software with something called makeup.
I've often wondered how "sucks" got to mean something bad.
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.
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?
How about noise and speech recognition? John lives in a house where speakers and microphones are place throughout the house. John is home alone and computer hears a loud noise. Computer "John are you alright?". Computer hears no response so computer gets John some help or computer hears John say yes and does nothing. John later decides to leave for 2 hours. When leaving the house John says leaving be back in two hours. Computer know house is empty so computer immediately reduces energy use. One hour later computer hears a noise. Computer ask what is the password. Computer hears either no response or incorrect password. Computer calls John on cell phone and lets John listen to noise. John than decides whether or not to call police. Computer hears running water but washing machine is not on so computer turns off water to that room. Computer hears smoke detector and hears the noise from fire so computer calls for help. I can think up a lot more problems a computer could help solve just by listening.
"Best pasta in town" Beth barista big town
No.. "highest rated pasta". Highest raped it pasta
No.. "great italian food" great stallion fooled
Fuckit.... typing it now...
I've often wondered how "sucks" got to mean something bad.
It's short for "sucks cock", which is basically another way of calling something gay.
Computer knows I'm socially awkward and instead offers to simulate a human female's speech while reciting lines from Star Trek.
Googled that for you:
http://www.worldwidewords.org/qa/qa-sou4.htm
Etymology dates to horse racing from the early 20th century, when horses would be injected with mysterious liquids ("soups") to improve their performance in races.
Well, the 90s are over too, and we have larger datasets now. With "large scale" SVMs still being measured in 10s of thousands of examples, you can see why companies with 4 orders of magnitude more *users* (let alone data items to classify) would need to use better scaling techniques. The older algorithms, when coupled with more modern minimizers, tend to fare well in comparison to the much smaller models you can train with more advanced techniques.
Also, as a researcher, you should recognize the adage about the actual order of importance for getting machine learning to work:
(1) picking the right features.
(2) getting enough data
(3) the learning algorithm
People love to talk at length about picking "the best" #3, when really you need to consider answers for #3 that let you do well on #2 and #1.
While I was a bit surprised to hear this Google project used networks (though not backprop trained NNs btw, which was the 80s fad), Andrew Ng is on the author list and he's a pretty smart guy (if you've done anything with reinforcement learning in the past 10 years you've probably run across his work). So I'm pretty sure they considered various options before they built something to run on 16K cpu cores.
You can read the ICML paper here:
http://research.google.com/pubs/pub38115.html