Generalization in the ANN

A network was created with 2000 records and aslo 5 networks were created with 400 records for each of them.(Note that the input data with 2000 records was divided to build 5 individual networks)Now the performance of each subnetworks is better than larg networks with 2000 records.Can we conclude that the larg network learnt too much from the examples given during the training, thus loosing the capability to generalize on the basis of new examples (overfitting)? but the small networks performed better because they had the less training records? Thanks a lot for any advice.

Answer 1

[Greg Heath]

> Can we conclude that the larger network learnt too much from the examples given during the training, thus loosing the capability to generalize on the basis of new examples (overfitting)? but the small networks performed better because they had the less training records?

ABSOLUTELY NOT!

Results heavily depend on how the data is divided. For example: randomly vs by sections.

You apparently misunderstand the concepts of overfitting and overtraining:

OVERFITTING: There are more unknown weights than training equations. This allows an infinite number of minima for training data ( How many solutions {x1,x2} are there for the problem x1+x2 = 1 ?!) which are not minima for nontraining (i.e., validation and testing) data.

OVERTRAINING: Training an overfit network beyond the point where performance on NONTRAINING data begins to deteriorate.

As long as all data is representative of the general I/O mapping, the more data, the better. That is why random datadivision is the default in MATLAB NN training programs.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Rita]

Thanks Greg for the clarification.I still can't understand why the dividing data (division by section 6 years of daily data) gives better results? since all data as you mentioned are " representative of the general I/O mapping".Why the network with almost 2000 records comparing to dividing them by years (365 records) did not perform well Thanks again for your precious comments

[Greg Heath]

No: IF the data are representative of the general I/O mapping.

This is easy to check:

Compare the performance of each net on all of the data.

[Rita]

Unfortunately, the performance of nets(5nets) with all of the data was not good.I also tried to examine the performance of each small nets with the data of each year and the results were not good too.so it seems that something wrong with data or nets?? Thanks again Greg