Fig. 2: Performance of fine-tuning and continual deep-learning system in various continual learning scenarios. | Nature Communications

Fig. 2: Performance of fine-tuning and continual deep-learning system in various continual learning scenarios.

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

Fig. 2

We present the performance of the (a) Fine-tuning strategy and (b) CLOPS in the Class-IL scenario which is characterized by six sequential tasks. We also present the performance of the (c) Fine-tuning strategy and (d) CLOPS in the Time-IL scenario which is characterized by three sequential tasks. The x-axis denotes the number of training epochs, the colored blocks reflect the task currently being trained on by the deep-learning system, and the y-axis denotes the average AUC. The performance of the fine-tuning system on tasks not currently being trained on degrades significantly, demonstrating catastrophic forgetting. CLOPS dramatically mitigates this catastrophic forgetting. The results are an average across five seeds and the shaded area represents one standard deviation from the mean.

Back to article page