Multi-task Learning without Forgetting [ECCV 2016] Zhizhong Li, Derek Hoiem. Train multi-task model in no old dataset. Distillation and fine-tuning. task-specific layers network expansion lower lr loss