Multi-task image classification on Multi-MNIST (test)
92.61Task 1 AccuracySoftmax MoE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Softmax MoEGate selection type=Baseline2021.06 | 92.61 | 91 | 8 | |
| DSelect-kGate selection type=Static2021.06 | 92.56 | 90.98 | 1.7 | |
| DSelect-kGate selection type=Per-example2021.06 | 92.42 | 90.7 | 1.5 | |
| Gumbel SoftmaxGate selection type=Static2021.06 | 92.3 | 90.63 | 1.8 | |
| Top-kGate selection type=Per-example2021.06 | 92.27 | 90.45 | 4 | |
| Top-kGate selection type=Static2021.06 | 91.93 | 90.03 | 4 | |
| Shared BottomGate selection type=Baseline2021.06 | 91.3 | 89.47 | — |