What Is Early Stopping?

Early stopping is a technique used in artificial intelligence (AI) or other computer learning programs in which the teaching temporarily stops in an attempt to improve scores. This can be done either through a series of modules or by interrupting a longer lesson several times. One problem that can occur from not using early stopping is that the AI memorizes information but does not learn. Another possible problem is that the AI continues to learn but loses information from other areas. This is a common feature in most AI systems that occurs automatically, but a technician may have to program this manually.

While most AI systems can learn from outside stimulation or through human interaction, a common way of teaching these systems before they are deployed or to supplement learning is through educational applications. These applications often teach new algorithms or new ways of solving problems. Early stopping can be used in two ways: the application can be split into modules and it stops after each module, or a long lesson may be interrupted by a stop.

If early stopping is not used, then the AI can suffer low test scores, showing it is not learning from the educational application. One way this manifests is through memorization. After a certain period — this differs for each AI system and teaching session — the AI system memorizes the information but does not understand it. This means memorized information can quickly be dropped, so this feature stops the learning process and forces the AI to display what it has learned.

The second problem that can occur without early stopping is more serious. Unlike memorization, this problem causes the entire AI to suffer and may be difficult to fix. In this scenario, the AI system will continue to learn from training, but this extra learning comes at the expense of other memory areas. It will begin dumping previously stored information to make room for new training. Early stopping keeps this from happening by allowing the AI to adjust its memory to better store new information.

This feature often is automatically used with most AI systems and training programs. If not, then a technician will have to manually run a stop at a certain point. When the AI is showing decreased test scores, a stop should be done immediately, because problems will appear after this point. While there are no serious problems with stopping earlier than this, it may impede the learning the process.