Multitask Learning on 8 vision tasks (accuracy)
92.4AccuracyFinetuned
Evaluation results
| Method | Links | |
|---|---|---|
| FinetunedBackbone=ViT-B/162025.02 | 92.4 | |
| FinetunedBackbone=ViT-B/322025.02 | 90.4 | |
| TIESBackbone=ViT-B/16, Basis size (M)=100%2025.02 | 79.7 | |
| L&SBackbone=ViT-B/32, Basis size (M)=100%2025.02 | 76.7 | |
| L&SBackbone=ViT-B/16, Basis size (M)=100%2025.02 | 75.9 | |
| TABackbone=ViT-B/16, Basis size (M)=100%2025.02 | 75.4 | |
| AE (Ours)Backbone=ViT-L/14, Basis size (M)=50%, Merging Method=TIES2025.02 | 74.2 | |
| TABackbone=ViT-B/32, Basis size (M)=100%2025.02 | 70.8 | |
| PretrainedBackbone=ViT-L/142025.02 | 69.8 | |
| AE (Ours)Backbone=ViT-B/32, Basis size (M)=50%, Merging Method=L&S2025.02 | 69.1 | |
| AE (Ours)Backbone=ViT-B/32, Basis size (M)=50%, Merging Method=TA2025.02 | 68.9 | |
| AE (Ours)Backbone=ViT-B/16, Basis size (M)=50%, Merging Method=TIES2025.02 | 67.2 | |
| AE (Ours)Backbone=ViT-B/16, Basis size (M)=50%, Merging Method=L&S2025.02 | 66.7 | |
| AE (Ours)Backbone=ViT-B/16, Basis size (M)=50%, Merging Method=TA2025.02 | 66.6 | |
| RandSelectBackbone=ViT-B/16, Basis size (M)=50%, Merging Method=TIES2025.02 | 66.4 | |
| RandSelectBackbone=ViT-B/16, Basis size (M)=50%, Merging Method=TA2025.02 | 64.5 | |
| PretrainedBackbone=ViT-B/162025.02 | 55.4 | |
| RandSelectBackbone=ViT-B/16, Basis size (M)=50%, Merging Method=L&S2025.02 | 55.3 | |
| PCABackbone=ViT-B/16, Basis size (M)=50%, Merging Method=L&S2025.02 | 52.3 | |
| PCABackbone=ViT-B/16, Basis size (M)=50%, Merging Method=TIES2025.02 | 49.6 | |
| PCABackbone=ViT-B/16, Basis size (M)=50%, Merging Method=TA2025.02 | 49.5 | |
| PretrainedBackbone=ViT-B/322025.02 | 48.1 |