My convolution neural network library

Finally, we've come to this. Today I present to you my convolution neural network library https://github.com/zen747/conv-neural-network. What's special about CNN is every layer is not composed of 1-dimensional neuron but 3-dimensional. Also, you have to implement the share-weights mechanism. It's much more complicate than bland neural network implementation. Still, the code should be easy enough to follow, check it out if you are interested.

The results seems decent enough.
Use a configuration as follows:

Input 32 32 3 0
Conv 32 3 1 1 0.1
Pool 2 2
Conv 64 3 1 1 0.1
Pool 2 2
Full 72 0.2
Full 72 0.2
Output 10
LearningRate 0.0004
RegularizationFactor 0.003
MaxNorm   1e4
BatchSize 150

I can obtain correct ratio about 70% after days of training. If my library supported GPU acceleration, it would take hours instead of days of training. But I will stop enhance the library at this point.