MLP Analysis and Related Systems
- This is a great system because it works pretty well
even when you don't know much about the data.
- Also, once it's learned, it's awfully fast to run. It's
just a few additions and multiplications and runs of
a sigma function.
- However, there are a lot of questions to ask in
developing one. How many layers? How many neurons?
- Overfitting is a problem. If you have relatively little
data, the system can, in essence, memorise it. It won't
then work on the test set (or real data).
- What has the system learned?
- The n-fold test (if done right) can give you an idea
of the power, but it's not statistical power.
- There are a lot of related systems. Radial Basis functions
are closely related.
- Support Vector Machines fit in.
- There has been a lot of work figuring these things out over the
last 30 years.
- It's not done.
