Brains “in the Clouds”Friday, June 29th, 2012 by Antonio Manzalini
Imagine we’re trying to build a system that can recognized pictures, for example of a cat. If we use standard machine learning, we have to collect thousands of labeled cat pictures to train the system. On the other hand, labeling takes a lot of time and work…
Recent advances on self-taught learning and deep learning are suggesting that we might be able to achieve this goal even using unlabeled data — such as random images fetched off the web or out of YouTube videos. These algorithms work by building artificial neural networks, loosely simulating brain learning processes.
It should be mentioned that neural networks are very computationally costly, so accuracy is limited as most networks used in machine learning have used up to 10 million connections. Obvioulsy, training much larger networks, might improve accuracy.
Interestingly, Google tested an artificial neural network spreading the computation across 16,000 of their data centres CPU cores, thus reaching more than 1 billion connections.
Then they show to this “small brain” YouTube videos for a week, without any image labeling. Neural network “discovered” what a cat looked like by itself from only unlabeled YouTube stills. That’s self-taught learning. More details in a Google+ post or in this full paper.
The research happened inside Google’s secretive X laboratory, where a small group of researchers began working several years ago on a simulation of the human brain, New York Times reports.
Same approach could probably be used to understand human speech, identifying complex data patterns or recognizing even a particular face of an individual…
In the future, down spiraling costs of storage and computing, and as such, the wide availability of huge clusters of machines in the cloud, will pave the way to these amazing applications, simulating brain learning and autonomic processes, just by using random data or images fetched off the web.