Permuted Pixel-by-Pixel Classification on Permuted MNIST (test)
96.7Accuracy (Clean)Dilated CNN
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Dilated CNN# layers=10, hidden / layer=50, # parameters (≈, k)=46, Max dilations=5122017.10 | 96.7 | — | — | — | — | — | — | |
| Dilated Vanilla# layers=9, hidden / layer=50, # parameters (≈, k)=44, Max dilations=2562017.10 | 96.1 | — | — | — | — | — | — | |
| CORNNhidden dimension (dh)=1282021.02 | 96.05 | 65.1 | 38.25 | 29.1 | 84.8 | 73.8 | 52.6 | |
| Lipschitz RNNhidden dimension (dh)=1282021.02 | 95.9 | 95.4 | 93.5 | 83.7 | 93.7 | 90.2 | 70.8 | |
| Zoneout# layers=1, hidden / layer=100, # parameters (≈, k)=42, Max dilations=12017.10 | 95.9 | — | — | — | — | — | — | |
| Dilated CNN# layers=10, hidden / layer=20, # parameters (≈, k)=7, Max dilations=5122017.10 | 95.7 | — | — | — | — | — | — | |
| Dilated Vanilla# layers=9, hidden / layer=20, # parameters (≈, k)=7, Max dilations=2562017.10 | 95.5 | — | — | — | — | — | — | |
| Dilated LSTM# layers=9, hidden / layer=50, # parameters (≈, k)=173, Max dilations=2562017.10 | 95.4 | — | — | — | — | — | — | |
| NRNNMultiplicative noise level=0.02, Additive noise level=0.02, hidden dimension (dh)=1282021.02 | 94.9 | 94.8 | 94.6 | 94.3 | 94 | 93.1 | 88.6 | |
| NRNNMultiplicative noise level=0.02, Additive noise level=0.05, hidden dimension (dh)=1282021.02 | 94.7 | 94.6 | 94.6 | 94.4 | 94 | 93.2 | 90.5 | |
| Dilated GRU# layers=9, hidden / layer=50, # parameters (≈, k)=130, Max dilations=2562017.10 | 94.6 | — | — | — | — | — | — | |
| Dilated GRU# layers=9, hidden / layer=20, # parameters (≈, k)=21, Max dilations=2562017.10 | 94.4 | — | — | — | — | — | — | |
| Dilated LSTM# layers=9, hidden / layer=20, # parameters (≈, k)=28, Max dilations=2562017.10 | 94.2 | — | — | — | — | — | — | |
| GRU# layers=1, hidden / layer=256, # parameters (≈, k)=200, Max dilations=12017.10 | 94.1 | — | — | — | — | — | — | |
| Full uRNN# layers=1, hidden / layer=512, # parameters (≈, k)=270, Max dilations=12017.10 | 94.1 | — | — | — | — | — | — | |
| Skipped RNN# layers=1, hidden / layer=95, # parameters (≈, k)=16, Max dilations=212017.10 | 94 | — | — | — | — | — | — | |
| Exponential RNNhidden dimension (dh)=1282021.02 | 93.3 | 90.6 | 78.4 | 61.6 | 80.4 | 70.6 | 51.6 | |
| Antisymmetric RNNbeta=1.0, gamma=0.001, learning rate=0.002, epsilon=0.01, epochs=100, hidden dimension (dh)=1282021.02 | 92.8 | 92.4 | 89.5 | 81.9 | 90.5 | 87.9 | 72.6 | |
| Skipped RNN# layers=9, hidden / layer=20, # parameters (≈, k)=11, Max dilations=2562017.10 | 91.8 | — | — | — | — | — | — | |
| LSTM# layers=1, hidden / layer=256, # parameters (≈, k)=270, Max dilations=12017.10 | 91.7 | — | — | — | — | — | — | |
| GRU# layers=9, hidden / layer=20, # parameters (≈, k)=21, Max dilations=12017.10 | 91.3 | — | — | — | — | — | — | |
| LSTM# layers=9, hidden / layer=20, # parameters (≈, k)=28, Max dilations=12017.10 | 89.5 | — | — | — | — | — | — | |
| Vanilla RNN# layers=9, hidden / layer=20, # parameters (≈, k)=7, Max dilations=12017.10 | 88.5 | — | — | — | — | — | — | |
| IRNN# layers=1, hidden / layer=100, # parameters (≈, k)=12, Max dilations=12017.10 | 82 | — | — | — | — | — | — | |
| Vanilla RNN# layers=1, hidden / layer=256, # parameters (≈, k)=68, Max dilations=12017.10 | 71.6 | — | — | — | — | — | — |