Does pre-training step help in convolution neural network?Posted: January 31, 2015
The answer is DEPENDS. When you have a lot of training data, it does not need to have pre-training step (unsupervised) because the weights eventually will converge to certain values. But if you don’t have much labeled data, then go ahead and do unsupervised pre-training.
See this: http://www.quora.com/When-does-unsupervised-pre-training-improve-classification-accuracy-for-a-deep-neural-network-When-does-it-not