This project explores Class-Incremental Learning (Class-IL), focusing on the challenge of catastrophic forgetting when sequentially adding new classes to a model using the FruitNet dataset which classifies five types of Indian fruits based on quality. The study empirically compares five different approaches implemented with a ResNet-18 backbone in PyTorch: simple Fine-Tuning, Joint training (as an upper bound), a basic Replay buffer method, Elastic Weight Consolidation (EWC), and Learning without Forgetting (LwF). The replay-based method demonstrated a good balance between learning new tasks and retaining old knowledge in these experiments.

You can find the complete code, and further details on the GitHub: