In calculation process, one system may select one of given data for each parameter tuning in a network. It is called Online Learning. This method is going to find appropriate parameters for the selected data. Therefore, the result may be far from expected values on all data.
The other system may select all data, calculate errors between each result and the corresponding teaching data, and tune parameters with regards to the average of errors. It is called Batch Learning. You may be able to find appropriate values, because all data are concerned on errors. When the given data is huge, it must be difficult for your PC to keep all error values from the data on the main memory.
Mini Batch Learning is to acquire strong points on both online learning and batch learning. The system selects some of all data randomly, and conducts the learning process. The number of data in one group is called Batch Size. The value is often from 10 to 100.