neural network
in [Matlab]
|
Prev: Fitting a (hyper)ellipsoid to data
Next: Dividing image to sub-blocks by dct2 and for each keeping only fixed set of coefficients.
From: gaurav goswami on 4 Jun 2010 06:31 hello,everybody I m working in the field of prediction through neural network . can any one suggest me that how the data should be divided for testing and training .Is there any pattern to divide the data or is it randomly 80% and 20% for training or testing . For my problem I have total 7x 225 data sets i.e. 75 for three ranges 0,0.1,0.2, now if i train the data for 0& 0.2 is it possible to predict for 0.1? Or should I train for all the 3 ranges after picking data randomly from every range.
From: Greg Heath on 5 Jun 2010 14:20
On Jun 4, 6:31 am, "gaurav goswami" <sagarsha...(a)gmail.com> wrote: > hello,everybody > > I m working in the field of prediction throughneuralnetwork . > can any one suggest me that how the data should be divided > for testing and training .Is there any pattern to divide the > data or is it randomly 80% and 20% for training or testing . > For my problem I have total 7x 225 data sets i.e. 75 for > three ranges 0,0.1,0.2, now if i train the data for 0& 0.2 > is it possible to predict for 0.1? > > Or should I train for all the 3 ranges after picking data > randomly from every range. The latter. I have covered these topics extensively in CANN (comp.ai.neural-nets) and CSSM. Searching those archives will yield many relevant posts. Choose Ntrn large enough to accurately estimate weights. If you are training to convergence with an I-H-O feedforward MLP with no jump-layer I/O connections (e.g., NEWFF/TRAINLM), then a good rule of thumb for Ntrn is Neq >> Nw where Neq = Ntrn*O (number of training equations) Nw = (I+1)*H+(H+1)*O (number of unknown weights) However, if regulation (TRAINBR) or Early Stopping with the val subset is used Ntrn can be somewhat lower. Choose Ntst large enough to precisely estimate errors i.e.., stdv(MSEtst) << MSEtst. The corresponding equations for standard deviations can be obtained from Wikipedia assuming MSE is distributed proportional to Chi-SQ and classification error counts are binomially distributed. Typically, choose Nval either Ntst/2 or Ntst. If N is large enough for the desired Ntrn + Nval + Ntst split, the percentages are determined. If not, consider M f-fold crossvalidation runs to decrease standard deviations by a factor of 1/sqrt(M*f). Search CANN and CSSM archives with keywords like greg heath Neq Nw greg heath stdv Ntst |