Lifelong Language Learning on Five Tasks (SQuAD, WikiSQL, SST, QA-SRL, WOZ) GPT-2 Base (test)
76.6Average ScoreMultitasked
Evaluation Results
| Method | Links | |
|---|---|---|
| MultitaskedDescription=All tasks are trained simultaneously2019.09 | 76.6 | |
| LAMOLSampling ratio (gamma)=0.2, Replay strategy=REAL2019.09 | 76 | |
| LAMOLSampling ratio (gamma)=0.05, Replay strategy=REAL2019.09 | 74.5 | |
| LAMOLSampling ratio (gamma)=0.1, Replay strategy=TASK2019.09 | 74.3 | |
| LAMOLSampling ratio (gamma)=0.1, Replay strategy=GEN2019.09 | 73.1 | |
| LAMOLSampling ratio (gamma)=0.05, Replay strategy=TASK2019.09 | 71.5 | |
| LAMOLSampling ratio (gamma)=0.05, Replay strategy=GEN2019.09 | 69.6 | |
| Fine-tunedDescription=Directly fine-tuned on the stream of tasks2019.09 | 51.5 | |
| MASMethod Type=Regularization-based2019.09 | 49.5 |