[baltcode]

First impressions:

1. It also covers "classical" artificial neural networks, i.e., things like backprop from before Hinton and others made breakthroughs for deep learning. This means you can start with this book even if you are new to ANNs. The later sections cover "real deep learning".

2. The language is great for beginners and users. You don't have to be an advanced math geek to follow everything. They seem to cover a fair amount of ground too, so its not dumbed down either.

3. I guess it covers most of the underlying theory and practical technicques but is implementation neutral. You should probably pick up a tutorial for your favorite implementation like Theano, TensorFlow, etc.

All in all, I like it a lot.

[ogrisel]

The old backprop from the 80's / 90's is still in use and the primary way to train deep nets. We tend to call it SGD (on a composition of differentiable functions) nowadays but it's the same algorithm.