Sequential Image Classification on MNIST Sequential (test)
99.48Accuracydense-IndRNN
Evaluation results
| Method | Links | |
|---|---|---|
| dense-IndRNN2019.10 | 99.48 | |
| coRNN# units=256, # params=134k2020.10 | 99.4 | |
| res-IndRNNNumber of layers=122019.10 | 99.39 | |
| IndRNNNumber of layers=122019.10 | 99.37 | |
| coRNN# units=128, # params=34k2020.10 | 99.3 | |
| R-Transformer2019.10 | 99.1 | |
| GRU# units=256, # params=200k2020.10 | 99.1 | |
| BN-LSTMimplementation=ours, hidden units=1002016.03 | 99 | |
| IndRNNNumber of layers=62019.10 | 99 | |
| DTRIV∞# units=512, # params=137k2020.10 | 99 | |
| LSTMimplementation=ours, hidden units=1002016.03 | 98.9 | |
| LSTM# units=256, # params=270k2020.10 | 98.9 | |
| FastGRNN# units=128, # params=18k2020.10 | 98.7 | |
| LSTM2016.02 | 98.2 | |
| LSTM2019.10 | 98.2 | |
| Transformerreported in=[26]2019.10 | 98.2 | |
| stanhs=21, l=12016.02 | 98.1 | |
| sTANH-RNN2016.03 | 98.1 | |
| anti.sym. RNN# units=128, # params=10k2020.10 | 98 | |
| iRNN2016.02 | 97 | |
| iRNN2016.03 | 97 | |
| RNN-path2019.10 | 96.9 | |
| uRNN2016.02 | 95.1 | |
| uRNN2016.03 | 95.1 | |
| uRNN2019.10 | 95.1 | |
| uRNN# units=512, # params=9k2020.10 | 95.1 | |
| IRNN2019.10 | 95 | |
| Best CVX-variantDepth L=3, Timestep=T = 22026.05 | 92.5 | |
| Best SG-variantDepth L=3, Timestep=T = 22026.05 | 91.3 | |
| stanhs=212016.02 | 87.8 | |
| LSTMs=32016.02 | 87.2 | |
| LSTMs=52016.02 | 86.4 | |
| LSTMs=72016.02 | 86.4 | |
| stanhs=132016.02 | 85.4 | |
| LSTMs=92016.02 | 84.8 | |
| tanh RNN Architecture (4)k=17, s=k2016.02 | 77.8 | |
| stanhs=92016.02 | 74.9 | |
| tanh RNN Architecture (4)k=21, s=k2016.02 | 71.8 | |
| tanh RNN Architecture (3)k=21, s=k/22016.02 | 69.6 | |
| LSTMs=12016.02 | 56.2 | |
| tanh RNN Architecture (3)k=17, s=k/22016.02 | 54.2 | |
| stanhs=52016.02 | 46.9 | |
| tanh RNN Architecture (2)k=21, s=12016.02 | 39.9 | |
| tanh RNN Architecture (1)k=17, s=12016.02 | 39.5 | |
| tanh RNN Architecture (1)k=21, s=12016.02 | 39.5 | |
| tanh RNN Architecture (2)k=17, s=12016.02 | 39.4 | |
| RNN(tanh)2016.02 | 35 | |
| TANH-RNN2016.03 | 35 | |
| stanhs=12016.02 | 34.9 | |
| Best CVX-variantDepth L=10, Timestep=T = 22026.05 | 27.1 | |
| Best SG-variantDepth L=10, Timestep=T = 22026.05 | 9.9 |