That is, finding observations about their data that are indicative of a prediction they are trying to make or a piece of data (say: a user or an image, perhaps a document) they are trying to label.
Deep Learning handles this for us. Typically machine learning takes a lot of domain knowledge. Deep Learning is capable of automatically finding correlations in the data and allowing us to focus on solving problems rather than worrying about what good indicators of data are.
Re: quick pretraining. Try again: deep learning pretraining takes forever to train due to how many passes through the dataset you need to make in order for it to make good approximations of the data.
That being said: deep learning is very receptive to map reduce type algorithms and online learning (streaming of data in mini batches)
That's why bigger companies tend to use it at scale.
That being said: it's astonishingly accurate despite being hard to train.
I will be more than glad to answer specific questions if you'd like.
I'll give you a few use cases:
Google uses it in their voice recognition on android.
IBM uses it in their Watson technology.
Google uses it in their Google+ image search.
It's being used for sentiment analysis at stanford.
I'm the author of a newer deep learning framework called deeplearning4j .
My goal with this is to make it easier to train these neural networks by just say: feeding it images and all you need to do is specify the labels: or text.
I'm also a collaborator/committer on another effort called metronome which is deeplearning on hadoop .
If you'd like an example of what deep learning is capable of, please take a look at . It's a very different way of doing machine learning.