Poor neural network performance can be described as two situations:
1. The error does not reach the goal.
SOLUTIONS:
> Raise the error goal of the training function.
> Use a different performance function.
> Increase the number of epochs for training in some situations.
2. The next issue that arises in neural network training is the speed and memory usage of training a network to reach the goal.
SOLUTIONS:
> Preprocess your data to make the network training more efficient.
> Try different training functions.
> Reduce the number of neurons that are being used. A general rule is to have less input paramters than output parameters.
Have a read of the attached file for more ideas.
I applied some heuristics (similar mentioned above) here in my Character recognition work:
http://www.google.com/url?sa=t&rct=j&q=naser%2Bgolam%2Bsarowar%2BIJCSNS&source=web&cd=1&ved=0CCYQFjAA&url=http%3A%2F%2Fpaper.ijcsns.org%2F07_book%2F200906%2F20090622.pdf&ei=WM_DTqDXHeWriAe3qoXjDQ&usg=AFQjCNHDnVNDUIE4woqjuvcacMD15r7rOQ&sig2=cbMkpA3hUo5e-AyOtbQB_A
If it's a supervised learning, is so important the performance of patterns for training. This is, if you don´t have a good set of patterns for training, perhaps never could be possible that ANNs learn. It is important consider too, differentes methods for learning process, differents algorithms: GA, MinMax, gradient descent...... I recomend one text: Neuronal Network for Pattern Recogntions by Christopher Bishop.
And, thank you very much to Mohammand for his paper. It's so interesting.
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Artificial Intelligence