Embodiments herein achieve a machine-learning system for managing shuffling of input training datasets. The machine-learning system includes a training dataset manager configured to shuffle an input dataset received from each of a plurality of electronic devices. Further, the training dataset manager is configured to split the input training datasets into a plurality of mini-batches. Each of the mini-batches, along with the target values, defines an error surface corresponding to an error function. A learning manager is configured to obtain a cross mini-batch discriminator based on the error function for each of the mini-batches. Further, the learning manager is configured to select a mini-batch associated with a least cross mini-batch discriminator from the plurality of mini-batches as optimal mini-batch.