The goal of this project is to make an MLP that can predict hand written digits correctly. The dataset is provided here.
So far I've added another hidden layer to the xor network, now both have 16 nodes. With that, I can start working on the MNIST data set with this network.
After breaking the model up into sub problems and tackling those, I have a neural network that will now classify hand written digits correctly ~96% of the time.
Some of the challenges I faced:
- Overflow error on the exponential inside the sigmoid.
- Changed the range of my values from 0-255 to 0-1, proportionally.
- Getting the correct shape for my weights.
- Understanding the notation for weights and applying that notation.
Thank you Michael Nielsen for writting the Neural Networks and Deep Learning online book, this really helped me understand neural networks and break down my network into sub problems.