What is the information storage capacity of a neural network? Firstly, a (feed forward) neural network can be thought as a lookup table. $f(x) = y$. So it is certainly storing some data.
The theoretical limit of data compression is -$\sum$ p $ln(p)$. The data in a neural network however is stored as weights and biases and not variable length coding, so does this equation even apply?
Say, I have a fully compressed data set of a certain size, will the size of weights and biases be the same if I have perfectly trained the network?
How do I know if my neural net has enough representative capacity or not to represent the entire data set?
 A: Well, a table is definitely not the right way to look at this. Consider the function $f(x)=x^2$. I can create an infinite table of input-output pairs that are represented by this function. However I can only represent exactly 1 such table with this function. Now consider this function: $g(x)=c\cdot x^2$. For different values of $c$, this function can represent an infinite number of tables (even an uncountable number).
The best way that I can come up with to describe the information storage capacity of a neural network is to quote the universal approximation theorem: https://en.wikipedia.org/wiki/Universal_approximation_theorem. To summarize it, say we have an arbitrary continuous function and we want to approximate the output of this function. Now say that for every input, the output of our approximation shouldn't deviate more than some given $\epsilon>0$. Then we can create a neural network with a single hidden layer that satisfies this constraint, no matter the continuous function, no matter how small the error tolerance. The only requirement is that the amount of nodes in the hidden layer might grow arbitrarily large if we choose the error-rate smaller and smaller.
A: 4 years later I have yet to see a concrete answer for this. The best I could find is this paper.
As @dimpol pointed out, it is useful to think of the neural network as a function with a finite number of parameters. If the number of parameters and the dataset match exactly then the function (neural network) is perfectly over fitted. This is the "storage capacity" so to speak.
If we go beyond that, something magical happens. It starts to generalize again. Its yet not fully understood why it happens.
https://arxiv.org/abs/1812.11118

Open AI also talks about it
https://openai.com/blog/deep-double-descent/

Edit (27/04/2022)
Almost 5.5 years later we still dont really know the answer but for large language models, we can estimate the optimal amount of training data that we should use to train the model.
Here is the paper from deepmind https://arxiv.org/abs/2203.15556#

