Agreed about scikit-learn. Anecdotally I've found that most "deep learning" and "big data" problems I see in industry could be trivially solved using scikit-learn on a reasonably powerful laptop.

Also if you want to really understand neural networks I can recommend Andrew Ng's new deep learning course on Coursea. You start with deriving the underlying math and implementing a simple (but fully functional) neural network using just basic numpy, so that when you get to tensor flow and the latest cutting edge techniques you actually understand what's happening under the hood and how things actually work.