Continual Learning on Split CIFAR-10

Summary

Comparison of anti-forgetting methods : Fine-Tuning, Experience Replay, Elastic Weight Consolidation (EWC), and Learning without Forgetting (LwF), with evaluation of backward transfer (BWT), forward transfer (FWT), and average accuracy across 5 sequential class-incremental tasks on Split CIFAR-10.

Dataset

CIFAR-10 is split into 5 sequential binary classification tasks, forming the Split CIFAR-10 continual learning benchmark.

Task	Classes
T1	airplane vs automobile
T2	bird vs cat
T3	deer vs dog
T4	frog vs horse
T5	ship vs truck

Representative sample images from each of the five tasks.

Contents

• continual_learning_split_cifar10.ipynb : Main notebook with implementation and experiments
• report_continual_learning_split_cifar10.pdf : Project report
• results.pt : Saved experimental results (accuracy matrices, metrics, etc.)
• figures/ : Saved visualization plots and figures

---

Made by: ABAZIZ Sarah & SEBBAH Sarah Farah
Course: "Advanced Machine Learning", Higher National School of Computer Science (ESI), Algiers