A Simple Machine Unlearning Approach Using Elastic Weight Consolidation

Machine Learning (ML) models have over the years been proved to be powerful tools for learning and inferring from data, but having the model unlearn a part of the data has been known to be a notoriously difficult problem. Machine Unlearning is an emerging field that deals with this exact problem - Having the model forget a subset of data without retraining from scratch. To accomplish this, we propose a method that uses Elastic Weight Consolidation (EWC) along with a modified loss term and show that the model is able to forget a subset of data while still retaining most information about the rest of the data. We evaluate this approach on the MNIST dataset using metrics such as Entropy, Confidence, and Accuracy.