If there is randomness in some part of the training process (and most surely there will be), then it makes sense. Different trainings may end up in different lo

If there is randomness in some part of the training process (and most surely there will be), then it makes sense. Different trainings may end up in different local minimums/maximums, so there is value in running multiple trainings simultaneously in different machines.