Neural Network
in [Matlab]
|
From: vitlap Esser on 21 Jul 2010 04:21 Dear all, I’m experiencing some problems and was hoping that you could help me? Firstly, whenever I try to create a Neural network of 140 input vectors and 36 target vectors, with 50 neurons in the hidden layer, I always get an error message from the training function regarding the memory. Do you maybe know what to do? Secondly, is there a way of training the network with only one vector and one target, instead of training it with the whole training matrix? Thirdly, I’ve tried to create a training set with noise as explained on your website, but that doesn’t work. Could you maybe tell me how the target matrix is supposed to look like? I haven't found a solution to this on your website, but if it's there, I'd be grateful if you could simply send me the link. Your help is very much appreciated. Please feel free to contact me if any of the above descriptions is unclear or if you need more information.
From: Greg Heath on 22 Jul 2010 00:20 On Jul 21, 4:21 am, "vitlap Esser" <v.es...(a)gmx.net> wrote: > Dear all, > I’m experiencing some problems and was hoping that you could help me? > > Firstly, whenever I try to create a Neural network of 140 input >vectors and 36 target vectors, with 50 neurons in the hidden layer, >I always get an error message from the training function regarding >the memory. Do you maybe know what to do? Who is "you"? The number of input vectors and target vectors must be the same. Therefore I am guessing that you are using incorrect terminology and 140 and 36 are the dimensionalities of the input and output vectors respectively. If you have Ntrn input/output pairs for training, what is Ntrn? For the design of an I-H-O = 140-50-36 MLP using NEWFF, you must estimate Nw unknown weights/thresholds using Neq equations where Nw = (I+1)*H+(H+1)*O = 141*50+51*36 = 8886 and Neq = Ntrn*O = Ntrn*36 For training to convergence it is desired that Neq >> Nw, i.e., that Ntrn >> ceil(8886/36) = 2469 What is your value of Ntrn? What are your values of Nval and Ntst? > Secondly, is there a way of training the network with >only one vector and one target, instead of training it >with the whole training matrix? If you mean sequential training where weight updates are made after the presentation of each input vector, the answer is yes. help adapt However, I don't think this will solve your problem. Why do you think you need I=140 input variables and O=36 output variables? If your input and target matrices are p and t, respectively, you need, at most, I = rank(p) and O = rank(t). Which variables to choose is another problem which should be dealt with in another thread. Why do you think you need 50 hidden nodes? I suggest using trial and error for finding the smallest H that will minimize MSE subject to the condition Neq >> Nw. > Thirdly, I’ve tried to create a training set with > noise as explained on your website, but that doesn’t > work. I didn't realize I had a website. You have to be more descriptive: How doesn't it work? >Could you maybe tell me how the target matrix is supposed >to look like? >Typically, noise is only applied to the input matrix p. > I haven't found a solution to this on your website, > but if it's there, I'd be grateful if you could simply > send me the link. > > Your help is very much appreciated. Please feel free >to contact me if any of the above descriptions is >unclear or if you need more information. Hope this helps. Greg
|
Pages: 1 Prev: 3D Radiation Pattern Next: find value in partly empty cell array |