Questions: Neural Network Approximation Theory

4 questions to test your understanding

Score: 0 / 4
Question 1 Multiple Choice

The universal approximation theorem guarantees that a single hidden layer network can approximate any continuous function. Does this mean deep networks (multiple hidden layers) offer no theoretical advantage over wide shallow networks?

ACorrect — depth is purely a practical convenience with no theoretical benefit
BNo — while shallow networks can approximate any function, they may require exponentially many neurons to do so, whereas deep networks can represent the same functions with polynomially many neurons (depth-separation results)
CNo — shallow networks can only approximate continuous functions, while deep networks can approximate discontinuous functions
DCorrect — the only advantage of depth is faster training via backpropagation
Question 2 True / False

The universal approximation theorem applies to neural networks with any non-linear activation function.

TTrue
FFalse
Question 3 True / False

The universal approximation theorem guarantees that for any target function and accuracy epsilon, there exists a set of weights that achieves epsilon approximation. It does NOT guarantee that gradient descent can find these weights.

TTrue
FFalse
Question 4 Short Answer

Explain the distinction between approximation power (what functions a network CAN represent) and learning/generalization (what functions a network WILL learn from data), and why the universal approximation theorem addresses only the first.

Think about your answer, then reveal below.