concept identical result in category deep learning

This is an excerpt from Manning's book Deep Learning with Structured Data.
You may be asking yourself – for the experiments shown in section 6.7, how was it possible to get consistent results between training runs? By default, aspects of the training process, including the initial weights assigned to nodes in the network, are set randomly. You will see this if you run repeated experiments with the deep learning model with identical inputs and parameter settings – you will get different results (such as the validation accuracy per epoch) each time, even if the inputs are identical. If there is an intentional element of randomness in training a deep learning model then how can we control this so that we can do repeated, controlled experiments to assess the impact of making a very specific change, such as the callbacks introduced in the example in this section? The key is setting a fixed seed for the random number generator. The random number generator provides the random inputs for the training process (such as initial weights for the model) that result in different results between training runs. When you want to have identical results from multiple training runs you can explicitly set the seed for the random number generator. I’ve incorporated the key points of this article in the model training notebook so you can, if you wish, run multiple experiments with identical results.