Fig. 1: Four continual learning scenarios. | Nature Communications

Fig. 1: Four continual learning scenarios.

From: A clinical deep learning framework for continually learning from cardiac signals across diseases, time, modalities, and institutions

Fig. 1

A deep-learning system, fω, is sequentially exposed to tasks with potential data distribution shifts. In the Class-IL scenario, we present mutually exclusive pairs of classes to reflect novel diseases. In the Time-IL scenario, we present data collected at different times of the year to reflect temporal non-stationarity. In the Domain-IL scenario, we present data from different input modalities to reflect different physiological sensors. In the Institute-IL scenario, we present completely different datasets to reflect disparate healthcare institutions. The system in the Class-IL scenario has a classification head that is specific to the task, and is thus task-specific. In the remaining scenarios, the system is task-agnostic; it is not aware of the task identity of the data.

Back to article page