The Subset Extended Kalman Filter

Currently, deep learning models are developed in an offline train-then-deploy paradigm using past data; However, this does not reflect many realistic applications where the underlying data-generating physical systems are constantly evolving. Dynamical systems naturally change and may drift away from the conditions under which the model was trained, causing prediction or classification error to increase.

Kalman Filtering has been used to update first-principles models since the moon landing.