MATLAB: Is it necessary to initialized the wieghts for retraining in matlab with nntool

MATLABneural network

i am doing a speech recognition problem; using nntool; i have found most of time that during the first attempt of training; the output is not good; so i have to retrain.

but my queries is: should i need to reinitialized the weights for retrain.

What if i don't re-initialized and train again?; actually i found better result without re-initializing.

i want to know which approach is right and why?

Best Answer

Some creation functions like newff, newfit, newpr, fitnet, patternnet and feedforwardnet are self initializing with random weights.

It is not clear if you are becoming successful by starting again with a new set of random weights or are just increasing the number epochs.

More information is needed.

Related Solutions

MATLAB: What are the reasons for having different outputs after several runs

The short answer is that train(only if all weights are zero) or configure( anytime) assign weights depending on the state of the RNG.

If you are training in a loop over random initial weights. The best statistically unbiased choice for the best net is determined by the minimum mse of the validation set.

 rng(4151941)             % Initialize RNG with famous birthday
 for i = 1:Ntrials
   s{i} = rng;            %Save the ith state of the rng (may not need a cell) 
   net  = configure(net,x,t);
   [ net tr ] = train(net,x,t);
    mseval(i) = tr.best_vperf;  % Best mseval over all epochs of ith run
 end
 [ minmseval ibest ] = min(mseval);
 rng     = s{ibest};            % For repeating the best design
 bestnet = configure(net,x,t);
 bestIW0 = bestnet.IW
 bestb0  = bestnet.b
 bestLW0 = bestnet.LW
 [ bestnet tr ] = train(bestnet,x,t) % NO SEMICOLON TO REVEAL ALL DETAILS!!!

Hope this helps.

Thank you for formally accepting my answer

Greg

MATLAB: MATLAB gives me different value of output every time I train a neural network, why

The default data division and weight initialization are both random.

To reproduce a design you have to know the initial state of the RNG before it is both configured with initial weights and divided into training, validation and testing subsets.

When designing multiple nets in a double for loop (creation in the outer loop and training in the inner loop), you only have to initialize the RNG once: before the first loop. The RNG changes its state every time it is called. Therefore, for reproducibility, record the RNG state at the beginning of the inner loop.

Exactly when the RNG is called differs for the different generation of designs. For special cases of the obsolete NEWFF family (e.g., NEWFIT, NEWPR and NEWFF), weights are initialized when the nets are created. For special cases of the current FEEDFORWARDNET family, (e.g., FITNET, PATTERNNET and FEEDFORWARDNET), weights can be initialized explicitly by the CONFIGURE function. Otherwise, they will be automatically initialiized by the function TRAIN.

When I find out exactly where the data is divided, I will post in both the NEWSGROUP and ANSWERS.