How hidden units represent features in deep learning

I'm trying to build a deep neural network for mobile phone data. I've been doing different tutorials, but there are some basic concerning how the hidden units represents features that I really would like to have clarified.

Look at the following picture (from slide 31 from www.cs.stanford.edu/people/ang//slides/DeepLearning-Mar2013.pptx by Andrew Ng):

1. How do acquire the more complex feature images from the hidden units? I would really like to get 100 random features images from each of my layers in my neural network as I think this would help me greatly in understanding how the complexity of the features increase, and what effect the hyper-parameters have on this.

2. Do you specify as part of a hyper parameter how much information (number of pixels) can be stored in a hidden unit in a given layer? It seems that the first layer has a substantially lower resolution than the others.

3. When combining the features from layer 3 (object parts) to layer 4 (object models), how does the specific hidden unit know how to place the different features? I would guess it would know that the hidden unit should represent 0.3 x "mouth feature" and 0.6 x "eyebrow feature", but how does it place the eyebrow above the mouth?

Thanks. Any books or websites with easy-to-understand material on this would also be greatly appreciated!