Neural network with large data set and many nodes

I want to make a neural network for an input of 371000 inputs and 371000 outputs. For start, I made 2 arrays for each 100 elements. So now I have 2 arrays with 3710 values. I tried with 1000 nodes in hidden layer but regression and performance is still low. I tried also with to hidden layers with 1000 each and bad results. If I put more than 1000 nodes it get me out of memory. I tried also to increase memory heap and virtual memory. Neural network is like this: net=newff(minmax(a_teresm_rot_int2),[1000 1],{'tansig','purelin'},'trainscg'). Some advices? Thank you!

[rocha nimic]

this is the error Out of memory. Type HELP MEMORY for your options. Error in C:\Program Files\MATLAB\R2012b\toolbox\nnet\nnet\nnderivative\+nnMATLAB\bg.p>bg (line 126)

Error in nnMATLAB.perfsGrad (line 7) [gWB,Perfs,PerfN] = nnMATLAB.bg... Error in nnCalcLib/perfsGrad (line 142) [trainPerf,valPerf,testPerf,gWB,trainN,valN,testN] = ... Error in traincgp>train_network (line 202) [perf,vperf,tperf,gX,gradient] = calcLib.perfsGrad(calcNet); Error in traincgp (line 72) [out1,out2] = train_network(varargin{2:end}); Error in network/train (line 329) [calcNet,tr] = trainFcn('apply',net,rawData,calcLib,calcNet,tr);

Answer 1

[Greg Heath]

 [ I N ] = size(input)
 [ O N ] = size(target)

if

 net.divideFcn = 'dividetrain'; %(Ntrn/Nval/Ntst = N/0/0)
 Ntrneq  = Ntrn*O               % No. of training equations.

For an I-H-O MLP, the number of unknown weights to be estimated using the Ntrneq equations is

 Nw = (I+1)*H+(H+1)*O

If Ntrneq > Nw, then H <= Hub where

 Hub = -1 + ceil( (Ntrneq-O) / (I+O+1))

When the data contains noise and measurement errors, it is desired that Ntrneq >> Nw resulting in H << Hub. If contamination is not too severe, Hub/10 <= H <= Hub/2 or smaller is a reasonable range to look for a good value of H. The smaller, the better.

If higher values of H are necessary, validation stopping (Nval >>1) or regularization (net.trainFcn = 'trainbr') is recommended.

If validation stopping is used, use one of the other divide functions. The resulting default sizes are

 [Ntrn/N, Nval/N, Ntst/N [~ [0.7,0.15,0.15]

Bottom line: Starting with H = 1000 is much to high. I recommend trying H = 10:10:100 (10 random weight initializations each ==> 100 trial designs). If any design looks promising, test it on ALL of the data. Be sure to record the state of the random number generator before each design so you can recreate it if the design is a good candidate.

Hope this helps.

Q1: Does I = O = 1?

Q2: Are the 10 data sets chosen randomly? If not, why not?

Thank you for formally accepting my answer

Greg