FFNet & PatternList & Categories: Learn...
|
|
You can choose this command after selecting one PatternList, one Categories and one FFNet.
Settings
- Maximum number of epochs
- the maximum number of times that the complete PatternList dataset will be presented to the neural net.
- Tolerance of minimizer
- when the difference in costs between two successive learning cycles is smaller than this value, the minimization process will be stopped.
Cost function
- Minimum-squared-error:
- costs = ∑allPatterns ∑allOutputs (ok - dk)2, where
- ok : actual output of unit k
- dk : desired output of unit k
- Minimum-cross-entropy:
- costs = - ∑allPatterns ∑allOutputs (dk · ln ok + (1-dk) · ln (1-ok))
Algorithm
The minimization procedure is a variant of conjugate gradient minimization, see for example Press et al. (1992), chapter 10, or Nocedal & Wright (1999), chapter 5.
Links to this page
© djmw 20040511