In some cases, we don't have all of the data we need for training before we start our learning. Sometimes, we are waiting for new data to arrive, perhaps the data we have is too large to fit into memory, or we receive extra data after a prediction has been made. In cases like these, online learning is an option for training models over time.
Online learning is the incremental updating of a model as new data arrives. Algorithms that support online learning can be trained on one or a few samples at a time, and updated as new samples arrive. In contrast, algorithms that are not online require access to all of the data at once. The standard k-means algorithm is like this, as are most of the algorithms we have seen so far in this book.
Online versions of algorithms have a means to partially update their model with only a few samples. Neural networks are a standard example of an algorithm that works in an online fashion. As a new sample is given...