neural network fitting time to resolve

Dear MATLAB Gurus,

This is the first time to use MATLAB, (have experienced with SAS for nearly 20 years and for better for value solution, I decided to go with MATLAB.) Today I carried out from menu screen Neural Net Fitting. X(input) has 1.2 million rows and 475 columns. Y has 1.2 million and only two response. (1 is good and 0 is bad). I tried to optimise using Levenberg-Marquardt and performance is MSE and calculation is MEX.

My machine is i7 3.4Ghz and 16GB ram and 600GB SSD integrated. I still watching the calculation and 1 Epoch takes around 30 minues and if you set 1000 then it takes 30*1000=30000 minutes??

I just wonder if this is the right estimate to see the results and if any of you have better parameter setting etc, please advise me.

just attaching a picture of iterations for you to understand easier. Thank you, Kaz

Answer 1

[Greg Heath]

Your data has the dimensions

[ I N ] = size(input) % [475 1.6M ]

[ O N ] = size(target) % [ 2 1.6M ]

However, since

1. You rarely need more than 20 or 30 samples per dimension for training, choose Ntrn ~ 30*475

2. Default ratios are Ntst = Nval = 3*Ntrn/14

3. You only need

   N ~  10*Ntrn/7 ~ 300*475/7 ~ 20,400

4. Although you have enough data to design about 1.6M/20.4K ~ 80 nets, combining the outputs from 10 or 20 should suffice.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Greg Heath]

Of course you could also use dimensionality reduction techniques to reduce the dimensionality 475 before applying the above strategy.

Greg

[KAZ]

Hi Greg, Thanks very much for you insight. Actually, what I mistook was I chose neural fit not the pattern recognition. I tried patter recognition and found it converges fairly quickly.

And off course your insight is very useful too for the efficient model development. I am not very familiar with MATLAB yet and need to learn more characteristic of the tools. As you also mentioned, it is quite handy to reduce the dimension as well after knowing the data well but for the moment I just want to stick to the original data's output before using principal components etc etc.