Neural network algorithms are stochastic by design and the source of randomness can be fixed to make results reproducible. The random number generators can be seeded in NumPy and TensorFlow which would make most of the Keras code 100% reproducible. But, there are some cases of additional sources of randomness and we can use these ideas on how to seek them out and perhaps fix them too. Read more at: http://machinelearningmastery.com/reproducible-results-neural-networks-keras/