Back Propagation Neural Network

For the third installment in the Biologically Inspired Computation series, we take a look at a more complex version of neural networking (as compared to our last installment where we investigated a Hopfield Network).

Theory

Essentially, we are running a network that we can train to recognize data, or solve a problem. We do this by giving the network sample data, and for each set of sample data, running the network through one time. We calculate the error of this run, adjust the weight of the connections between all the neurons in the entire network, and repeat. After running through all the sample data we have adjusted the weights to ‘recognize’ the data. Then, given another set of data, the network is trained to perform the same operation, such as classification (maybe determining forged vs authentic bank notes) or just solving a mathematical equation, such as we do in this particular example.

Code

This simulator is written in Cocoa, requires Xcode 3.1 or later, and is Leopard only. Subclasses NSOperation for multithreading, and have two subclassed NSViews for a ‘Safari Downloads’ style window, just showing the progress of each running experiment, with a cancel button beside it.