Selective Unlearning of Training Data Without Complete Retraining

An NSF-sponsored research project on ensemble models for machine unlearning. Machine unlearning is the problem of forgetting a data’s contribution to a model to adhere to privacy laws and protect user data. Using the CIFAR-10 dataset, we created an ensemble of 5 ResNet-18 models tasked to forget some data. Compared to a larger ResNet-152 model, the ensemble model had better validation accuracy.