This is a multilayer perceptron implementation written in C++ with Eigen. Everything is vectorized with Eigen - there aren't any nested loops used to evaluate/train the network.
It supports any number of fully connected layers of any size. It trains using backpropagation. It supports RPROP as a weight update mechanism.
All of the important code is in src/MLP.cpp.
Provided you have a C++ compiler and Eigen headers installed, you should be able to just run make. Depending on your installation you make have to refer to your Eigen headers with -I, which can be specified using OPT in the makefile.
The binary will be in dist/cppnn. Right now main.cpp includes a basic example that does training of a small (not minimal) network to fit XOR. mnist.cpp (which main calls right now) trains on MNIST data. As it is, it gets about 90% accuracy after ten seconds or so of training. It needs the MNIST data in a directory you pass in as an argument.
- RPROP tweaks (need to limit weights, among other things)
- Parallelization
- MNIST preprocessing (width/brightness normalization, experiment with DCT)