fiddle2 entry

Date: 2004-04-07 01:05:12 (Author: trav)
Link: http://travis.kroh.net/archives/002936.php

---

From the documentation of Neuralyst:

...we dont' know exactly what characteristics the neural network has determined are relevant (this is particularly important with small sample sets...where there will be fewer or no samples to contradict bad generalizations) and there is often no way to find out without experimenting with the neural network. Theoretically it should be possible to understand the model developed by the neural network through a detailed examination and understanding of the weights, but in practice these values and relationships are often too complex for this to be attempted.

This is where I think a lot of data mining software effort should be focused.
It's probably due to my inexperience, but more interesting to me than using neural nets to learn things so we don't have to, is the idea of using neural nets to learn things so they can teach us what is important. Imagine the benefits this could reap for the communities of law enforcement, medicine, economics, etc... if we had a way to easily analyze the weights of neural networks. By mining the weights, and producing an intuitive visualization, we could more effectively do research. The machine wouldn't HAVE to learn to generalize, it would just need to learn what's important, so we could more easily do the generalization ourselves.