Fig. 3: Marginal benefit of storage and acquisition mechanisms on performance of continual deep-learning system. | Nature Communications

Fig. 3: Marginal benefit of storage and acquisition mechanisms on performance of continual deep-learning system.

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

Fig. 3

We show three different learning strategies in the Class-IL scenario. Random storage and acquisition stores instances into, and acquires them from, the buffer randomly. Random acquisition stores instances into the buffer using our importance-based strategy and acquires them from the buffer randomly. Random storage stores instances into the buffer randomly and acquires them from the buffer using our uncertainty-based strategy. The results are shown as a function of the storage fraction, b, and acquisition fraction, a, and are an average across five seeds. Improvement in performance of the random acquisition and random storage learning strategies relative to the random storage and acquisition strategy points to the benefit of our storage and acquisition mechanisms, respectively.

Back to article page