I just started learning about neural networks and was wondering what a neural network with 2 hidden layers is able to express over a neural network with just 1 hidden layer (where number of neurons are limited). Specifically, I am trying to come up with an binary classification example with a decision boundary that can be expressed with 2 hidden layers but not one (assuming input is a point in the 2D space, ReLU activation for the hidden layers, and sigmoid activation for the output). The 1-layer network has 4 hidden neurons in its layer while the 2-layer network has 2 neurons each layer, so still 4 hidden neurons total. I've tried looking at concentric circles, enclosed shapes, and nonlinear boundaries but have no luck so far.
What is an example of a binary decision boundary that 2 hidden layers (2 neurons each layer) are able to capture but 1 hidden layer (4 neurons) isn't? Is this even possible?
Update: I edited my question to a limited number of neurons as that the universal approximation theorem does not apply.
