University of Toronto Professor Geoffrey Hinton gave the second talk of the year as part of the Yahoo Labs Big Thinkers Series in Santa Clara on March 18th to discuss why the brain needs to learn by following the gradient of an objective function. He argues that learning a model of the world using unsupervised data might be useful as a first step. The motivation for this approach comes from behavior displayed by human/monkey brains, which tend to learn a model of the world using lots of unsupervised information. The way Hinton employed this idea was to use unlabeled data to learn some preliminary weights for the hidden units in his neural networks while optimizing modeling fidelity. Using labeled data, these weights were fine-tuned to obtain the best possible accuracy. The advantage is that, in most real world situations, the amount of unlabeled data available to algorithms is many orders of magnitude larger than the amount of labeled data. Using this method, Hinton’s team was able to get much better predictive accuracy on some standard applications like voice recognition.

More information at:

http://labs.yahoo.com/Big_Thinkers