Parameter Regularization

Learning without Forgetting (LwF)

Approach that uses knowledge distillation to preserve the model's responses on old data while learning a new task. The original model serves as a teacher to guide the updated model, thus avoiding performance degradation on previous tasks.

← Kembali