In this project I accomplished two goals:
Without using tools from MATLAB's neural net toolbox, I created a two-layer neural net that computes a piecewise linear approximation of an input function f.
- layer 1 uses the ReLU nonlinearity
- layer 2 does not use nonlinearity since it uses the ReLU as the identity
A small neural net was trained on data with binary labels. Network structure and training parameters were tuned until achieving 97.12% testing accuracy.