Online Learning
lastmod 2021-02-19created 2021-02-191 edits
In online learning, the system is trained by feeding it data sequentially, either one instance at a time or in small groups called mini-batches.
Each training step is fast and cheap, so the system can learn from data immediately as it arrives. Online learning suits systems that receive data continuously and must adapt to rapid change on their own. It is also suitable when computing resources are limited. Once data has been learned from, it is no longer needed and can be discarded.
One important parameter of an online learning system is how quickly it should adapt to changing data. This is called the learning rate. With a high learning rate, the system adapts to data quickly but also quickly forgets old data. Conversely, with a low learning rate, the system has more inertia and learns more slowly. However, it becomes less sensitive to noise in new data and to unrepresentative data points.
The biggest problem with online learning is that system performance can gradually decline when bad data is fed into the system. To reduce this risk, the system must be monitored closely, and training must be stopped immediately when a drop in performance is detected.