Why does using the parallelization option slow down my neural network simulations?

Hi,

I am using the neural network toolbox to train and simulate a feedforeward neural network in the following manner:

net1 = feedforwardnet(60); net2 = train(net1,inputs,targets,'useParallel','yes') outputs = net2(input_big,'useParallel','yes');

Theese comands are taken straigth from mathworks homepages: http://www.mathworks.se/help/nnet/gs/parallel-computing.html

Without parallelization the code takes about 26 seconds to finish, with parallelization it varies somewhat, but takes between 40-60 seconds. I am using a pool of four workers on an intel core i7 with R2013a.

If you have an idea for something I could try to speed things up it would very appreciated.

Pål Egil

[Mark Hudson Beale]

What size dataset are you using? How much RAM do you have? Are you setting the random seed so that each run is comparable? (Different random initial conditions can result in different times.)

Here are two runs to compare with:

------

For the following code, I get 87 and 40 seconds in R2013a with 4 workers.

[x,t] = vinyl_dataset;

net1 = feedforwardnet(10);

rng(0), tic, net2 = train(net1,x,t); toc

matlabpool open 4

rng(0), tic, net3 = train(net1,x,t,'useParallel','yes'); toc

-------

This code ran in 41 and 10 seconds:

[x,t] = house_dataset;

net1 = feedforwardnet(60);

rng(0), tic, net2 = train(net1,x,t); toc

rng(0), tic, net3 = train(net1,x,t,'useParallel','yes'); toc

------

What times do you get?

Mark

[Giovanni Barbarossa]

I have the same problem with patternnet. The training is 5 times slower with the parallel option. It makes no sense.