One of the neurons in the artificial neural network, trained from still… (Google )
Google researchers and Stanford scientists have discovered that if you show a large enough computing system millions of images from random YouTube videos for three days, the computer will teach itself to recognize ... cats.
That may sound inconsequential at best and downright ridiculous at worst -- but in fact, it is very important.
The research shows that if a computer is big enough, and programmed correctly, it can learn to make sense of random, unlabeled data, in just days without any help from humans.
And this research is especially important to Google because it has major implications for search.
As of now, Google is very good at searching labeled data, such as cat images that are labeled as cat images. But imagine the improvement to search if Google's computers were able to find cat images that aren't labeled cat images, for example. Or even if Google could "see" a cat in the real world and recognize what it is, and then tell you information about it. After all, there is much more unstructured data floating around our world than labeled data from a computer's perspective.
It is possible, of course, to teach a computer what a cat looks like. You can show the computer thousands of pictures of cats so that it learns what constitutes a cat's face -- pointy ears, almond eyes, furry face.
But then if you want the computer to learn what a crocodile looks like, you'd have to show it thousands of pictures of crocodiles, too.
It is a time-consuming and expensive process.
For this paper, the researchers worked with the idea of "deep learning," which involves software that loosely simulates the learning process of the human brain, building a "neural" network with 1 billion connections.
The basic idea is that you throw a ton of information at the computer and wait for it to start sorting that information on its own. So instead of searching for 10,000 pictures of cats, you can just show the computer millions of YouTube videos -- which will obviously include thousands of images of cats -- and the computer will come up with the concept of a cat on its own.
Andrew Y. Ng, a computer scientist at Stanford University who lead the research with Google fellow Jeff Dean, says in some ways it is analogous to what would happen if you showed an infant nothing but YouTube videos for a week. The baby would not be able to tell you it saw a cat of course, but the idea of a cat would have imprinted itself on the baby's brain.
For the record, the computer did not only learn to recognize cats from this experiment, it also learned to recognize human faces, and body parts and probably other images as well.
This is not the first time that deep learning ideas have been tested, but it is the biggest computer system that has ever tested them, and that has led to a whopping 70% increase in accuracy in standard image classification tests compared with similar tests run on smaller systems.
"To make the analogy, the bigger the simulated brain, the smarter you can be," Ng said.
For this paper, the researchers worked with 16,000 CPUT cores and created a model with more than 1 billion connections.
That's impressive, but what we've got in our heads is even more impressive. The researchers say an adult human brain has 100 trillion connections.
Ng said his next step is to experiment with even larger computer systems.
"We found pretty consistently that the bigger the simulated brain we can build, the more interesting things we can learn," he said. "We are focusing a lot of our efforts on technology to scale it up even bigger and build a really massive network."
18-year-old goes to school by day, gets Apple scoops at night
Analysts expect just 10-cent rise above IPO price for Facebook
Lying about your weight online, no longer a crime in Rhode Island