The most urgent targets are:
|
- make the backprop package fully functional in console mode.
- finish the cascade-correlation net and publish it on the net as well.
- make a graphical interface (Qt, I think would be best, but GTK+ or even better,
GTK- is also an option) that would include tools for analyzing the status of the net,
telling what is wrong etc. It would let the user "touch" the weights, the units
and let him have complete control over the net's training.
- the net is written in C++. Even a simple neuron is an object. That provides the program
with flexibility but is not so kind to its speed and memory needs. A very nice feature of
the program should be, thus, its ability to dump simple C code which should
simulate very quickly the given net.
|
|