Dynamic Mitigation of Catastrophic Forgetting Using the Sampling Network

The catastrophic forgetting in transfer learning makes a neural network lose the performance on previously learned datasets when dealing with large amounts of data. Predictive Elastic Weight Consolidation (PEWC) reduces the catastrophic forgetting by extracting only images with relatively more incorrect network predictions, but uses static sampling technique. PEWC also includes images in the training data which can be correctly classified by the network, leaving the possibility for further reduction of the training data. In this paper, we additionally apply a sampling network that extracts images dynamically without sorting, so that only images whose predictions are similarly inaccurate in general are used for training. In the experiment, our method achieved a similar level of mitigation of catastrophic forgetting while learning less data than PEWC.