Today (13 Sep 2016) I happened to find my 2003-2004 experiments with neural networks and specifically image compression. The GPU times were very far and the interest in neural networks was relatively low at the time, however I really loved the concept and used some free time to play with them.
With my surprise, the code still compiles and run! Compressing a few test images using 8x8 blocks as inputs, an interediate layer of 8 neurons, and an 8x8 output.
The neural network implementation uses back propagation and has a few configurable parameters. I guess it's a bit different compared to the new NN implementations. For instance the sigmoid function is used, and only CPU training is supported, of course.
I can't really recall the details, but I'm pushing this now completely useless code just for archiving goals.