Extended Data Table 2 Application of FIP to SplitCIFAR continual learning with convolutional neural network architectures

From: Engineering flexible machine learning systems by traversing functionally invariant paths

  1. Accuracy of FIP on 20 task SplitCIFAR for CNN architectures and comparison to Elastic Weight Consolidation (EWC), Relevance Mapping Networks (RMN), Gradient Episodic Memory (GEM), Brain inspired generative replay (BR).Results for RMN, EWC, GEM, and BR are compiled from the literature. Bold font used to highlight that FIP has highest accuracy in continual learning task across methods.
Back to article page