Continual Learning with Bayesian Model Based on a Fixed Pre-trained Feature Extractor

Current deep learning models are characterised by catastrophic forgetting of old knowledge when new classes. This poses a challenge in intelligent diagnosis systems where initially only training data limited number diseases available. In this case, updating the system with would inevitably downgrade its performance on previously learned diseases. Inspired process human brains, we propose Bayesian generative model for continual built fixed pre-trained feature extractor. model, each class can be compactly represented collection statistical distributions, e.g. Gaussian mixture models, and naturally kept from learning. Experiments two skin image sets showed that proposed approach outperforms state-of-the-art approaches which even keep some images classes during