Given a neural network
bounding the error as a function of the number of trainable parameters,
For my master's thesis, I proved that a rate of up to
For target functions of varying smoothness we use, a Gaussian, a cusp and a step:
The neural network
The main result of these experiments is summarized by the following figure. A rate
A crucial point in constructing our proof arguments is that as
Each
See: here


