Continual Distillation Learning for Rehearsal-Free Class-Incremental Learning via Decoupled Prompting
The episode discusses a paper on Continual Distillation Learning for Rehearsal-Free Class-Incremental Learning via Decoupled Prompting. The hosts analyze why standard distillation methods fail in this setting due to 'distillation information forgetting' and introduce 'KD prompts' as a solution. Experiments show KD prompts significantly improve accuracy and reduce forgetting rates across different continual learning methods.