Predictive EWC: mitigating catastrophic forgetting of neural network through pre-prediction of learning data

초록

Each time an artificial neural network learns an unseen dataset, it loses its ability to recognize the feature that it had learned before. This phenomenon is called the catastrophic forgetting problem (CFP). In image classification, the representative feature of each class that has significantly contributed to determining the class into which a given an image is categorized and thus directly influences performance. CFP can thus be damaging. The proposed algorithm, called Predictive EWC or PEWC, learns only sampled data from a new task consisting of the most challenging images for the network to classify. The criterion for extracting a sample is the absolute value of the difference between the network’s predicted value and the annotated value of the given image. This reduces the size of the task to be learned and mitigates the likelihood of CFP. An experiment showed that the average accuracy of a given task is 5% higher when the proposed algorithm is used in comparison with a prevalent algorithm, EWC, while consuming fewer resources. © 2019, Springer-Verlag GmbH Germany, part of Springer Nature.

키워드