Neural network training in Azure and Weka

Question

I'm using some data-sets (.arff files) to train neural networks in Weka and Azure. For every data-set , I get an overall accuracy difference between WEKA/Azure up to 5%. Obviously, I use the same training parameters, like number of iterations, learning rate, momentum, etc. Is this difference justified?

Answer 1

Of course, if your input data files are similar, still contain slightly different data values, then classification results are unlikely to be exactly the same between runs, and 5% variation is not unheard of.

Perform some experiments with a single dataset and only vary the "seed" attribute (maybe try 5 differnt integers), that will tell you how much randomness is involved in running your algorithm, MultilayerPerceptron I suppose.

(Seriously, what kind of answer do you expect here?)