::: Backprop c++ code :::

The network lets you create networks of arbitrary size, and uses a sigmoid function to compute the output of each unit, weights and biases are adjusted with the delta rule as described by the PDP researchers . The net has three layers which are fully interconnected. It uses on-line learning. I successfully trained this network to do character and number recognition.

Download C++ source code of my backpropagation neural network
Download a sample data file for the net that computes 4 logic operators from two given inputs

I compiled the source code on Borland turbo C++ version 3, and on Microsoft visual C++ version 5. I ran the program (it uses a DOS interface) on DOS, windows 95, and windows NT.

How to use it

Run the program, then load the data file you can download from this page, or choose the option "make data file ". The program asks you to specify values, a nice learning rate to start with is 0.5 . Use binary values for the patterns. Weights and biases are given random values automatically. Then load the file you just created with the option "load data file". Then select option "learn from data file", in this way the network will be trained with the data you specified. When the network is trained successfully, test it with the "compute output" option. Don't forget to save your results before exiting! If the training seems to don't give any results, specify a lower learning rate (the training process got stuck in a local maximum).

[index] [mail me]
Thomas Riga, University of Genoa, Italy