This project has two classes that are important for doing stuff - Classifier and Trainer. Both accept a
reference to the Network.
The Network is a container for the data model. It has internal objects like Neurons and Synapses
that do their own logic internally. The network creates layers of these smaller objects when it is
initialised.
The Classifier and Trainer classes manipulate the Network.
The ./bin/train program will use the Classifier and Trainer classes, while ./bin/classify will only
initialise and use the Classifier object.
I found the biological model of a neural network to be useful to some extent when doing this project. Here are the original pics I made to help me think about it:
The Soma concept above I got from studying biological networks, but with this program, those functions
are not internal to Neurons and the Classifier handles everything to do with forward feed.
The concept of backpropagation deserves an explanation in itself. I've found Matt Mazur's explainer to be great, if you're prepared to step through the example step by step. Videos also helped my understanding of it.
I've chosen to use a logistic function for activating the neurons. See the Classifier functions for
details on that. Two other functions look useful. However, the backpropagation will need to change
if the logistic function is not used because the backpropagation algorithm I use requires
partial derivatives of the logistic function.
