Updating the model continuously as new data arrives, rather than retraining in batches. Suits fast-moving environments such as recommendation, pricing and fraud, where yesterday’s model is already stale. The cost is that the system can be moved by whatever arrives next, including deliberately.
Technical view
Incremental parameter updates on streaming data, evaluated by regret against the best fixed hypothesis in hindsight. Practical concerns include catastrophic forgetting, sensitivity to non-stationarity, and the difficulty of rollback once bad data has been absorbed. It also creates a poisoning surface: an adversary who can generate traffic can steer the model. Most production systems prefer frequent scheduled retraining with validation gates.
Related
Part of AI Terminology Index, section: Classical Machine Learning.