This package implements a very simple multilayer perceptron network in Go, with gradient descent training via backpropagation.

Included is a simple example to train a 6 node network on the XOR function. Of course, the example doesn't output anything really impressive, just iterates until the mean squared error of the output is below a certain threshold for all training examples. Doesn't take long at all on my laptop, though.

The other example is mnist.go, which is designed to train on the MNIST handwritten digits dataset. I've got it up to over 96% accuracy in a few epochs, like so:

$ wget -r -np -Agz http://yann.lecun.com/exdb/mnist/
$ mv yann.lecun/exdb/mnist/*.gz .
$ for i in *gz; do gunzip $i; done
$ ./mnist -si train-images-idx3-ubyte \
          -sl train-labels-idx1-ubyte \
          -ti t10k-images-idx3-ubyte \
          -tl t10k-labels-idx1-ubyte

I'm sure this codebase could easily be made to do better. See http://yann.lecun.com/exdb/mnist/ for more details on the dataset.

Mostly, I did this to experiment with building things in Go, and because I'd never actually successfully implemented backpropagation before. This code is Public Domain; do what you like with it. It is not guaranteed to work or to be useful for any purpose. Patches welcome!

SDE

San Francisco, CA

Written 2011/10/16, updated 2013/10/04