Small experiments with artificial neural networks in Torch7.
Learn with the delta rule how to classify linear seperable data, the script will animate the separation line during training.
Learn with backprop how to classify non-linaer seperable data.
A classical encoder problem where the network is forced to find a compact coding of sparse data. The network has a hour glass shaped topology (8-3-8).
Learn with backprop how to approximate a function (Gauss), the first plot show the exact function values and the second shows the approximation. Experiment by adjusting the number of hidden nodes and watch how the network will underfit or overfit.
KTH notes.