Image Classification on MNIST standard (test)
99.7AccuracyDF blockwise
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| DF blockwise# Layers=102025.01 | 99.7 | — | — | |
| ASGE# Layers=112025.01 | 99.59 | — | — | |
| MF cumulative*# Layers=42025.01 | 99.58 | — | — | |
| MF final*# Layers=42025.01 | 99.58 | — | — | |
| DF layerwise# Layers=102025.01 | 99.53 | — | — | |
| BP*# Layers=42025.01 | 99.52 | — | — | |
| DeeperForward# Layers=42025.01 | 99.5 | — | — | |
| FAUST# Layers=52025.01 | 99.5 | — | — | |
| Channel-wise FF# Layers=42025.01 | 99.42 | — | — | |
| LogicTreeNet-LModel=CNN, Gate count=1.27 M2026.02 | 99.35 | — | — | |
| PearsonNoise Type=Sparse, Noise Level=Low, Bias Correction=false2020.11 | 99.24 | — | — | |
| JeffreyNoise Type=Sparse, Noise Level=Low, Bias Correction=false2020.11 | 99.24 | — | — | |
| LogicTreeNet-MModel=CNN, Gate count=566 K2026.02 | 99.23 | — | — | |
| BitLogicModel=FFN, Gate count=384 K2026.02 | 99.15 | — | — | |
| SchrödingerHidden dimension=64, Depth=3, Batch size=16, Dropout=0.1, Global mean pooling=true, Epochs=200, Optimizer=Adam, Learning rate=3 × 10−4, Number of random seeds=52026.05 | 99.13 | — | — | |
| CNNBatch size=16, Dropout=0.1, Epochs=200, Optimizer=Adam, Learning rate=3 × 10−4, Number of random seeds=5, Overall width/depth scale=Matched to GNNs2026.05 | 99.07 | — | — | |
| CaFo# Layers=32025.01 | 99.04 | — | — | |
| FA*# Layers=42025.01 | 98.99 | — | — | |
| LILogicNet-LModel=FFN, Gate count=32 K2026.02 | 98.95 | — | — | |
| MPNNHidden dimension=64, Depth=3, Batch size=16, Dropout=0.1, Global mean pooling=true, Epochs=200, Optimizer=Adam, Learning rate=3 × 10−4, Number of random seeds=52026.05 | 98.95 | — | — | |
| DTP# Layers=62025.01 | 98.93 | — | — | |
| OursP(s=1)=0.5, Base loss=MAE, Backbone=CNN2026.05 | 98.86 | — | — | |
| SNN (rate coding)Network=784-800-10, Coding=SNN (rate coding), BP scheme=surrogate-gradient, Spike type=multi, Leakage=leaky, τV=τ, τI=τ2022.11 | 98.84 | — | — | |
| Trans-PRODENP(s=1)=0.5, Base loss=CCE, Backbone=CNN2026.05 | 98.73 | — | — | |
| OursP(s=1)=0.7, Base loss=MAE, Backbone=CNN2026.05 | 98.73 | — | — | |
| OursP(s=1)=0.3, Base loss=MAE, Backbone=CNN2026.05 | 98.71 | — | — | |
| SCFF# Layers=52025.01 | 98.7 | — | — | |
| JeffreyNoise Type=Random, Noise Level=0.7, Bias Correction=false2020.11 | 98.67 | — | — | |
| Trans-MCL (MAE)P(s=1)=0.5, Base loss=MAE, Backbone=CNN2026.05 | 98.66 | — | — | |
| Trans-PRODENP(s=1)=0.3, Base loss=CCE, Backbone=CNN2026.05 | 98.64 | — | — | |
| SNN (rate coding)Network=784-800-10, Coding=SNN (rate coding), BP scheme=surrogate-gradient, Spike type=multi, Leakage=leaky, τV=τ, τI=τ2022.11 | 98.64 | — | — | |
| Trans-PRODENP(s=1)=0.7, Base loss=CCE, Backbone=CNN2026.05 | 98.62 | — | — | |
| DMINoise Type=Sparse, Noise Level=Low, Bias Correction=false2020.11 | 98.59 | — | — | |
| DFA*# Layers=42025.01 | 98.53 | — | — | |
| DiffLogic Net (largest)Model=FFN, Gate count=384 K2026.02 | 98.47 | — | — | |
| LogicTreeNet-SModel=CNN, Gate count=147 K2026.02 | 98.46 | — | — | |
| LILogicNet-MModel=FFN, Gate count=8 K2026.02 | 98.45 | — | — | |
| timing-based backpropagation for multi-spike SNNsNetwork=784-400-10, Coding=SNN (temporal coding), BP scheme=timing-based, Spike type=multi, Leakage=leaky, τV=τ, τI=2τ2022.11 | 98.43 | — | — | |
| ANNNetwork=784-800-10 (+ dropout), Coding=ANN2022.11 | 98.4 | — | — | |
| GINHidden dimension=64, Depth=3, Batch size=16, Dropout=0.1, Global mean pooling=true, Epochs=200, Optimizer=Adam, Learning rate=3 × 10−4, Number of random seeds=52026.05 | 98.33 | — | — | |
| PEPITA# Layers=22025.01 | 98.29 | — | — | |
| Trans-MCL (EXP)P(s=1)=0.5, Base loss=EXP, Backbone=CNN2026.05 | 98.27 | — | — | |
| timing-based backpropagation for multi-spike SNNsNetwork=784-400-10, Coding=SNN (temporal coding), BP scheme=timing-based, Spike type=multi, Leakage=non-leaky*, τV=τ, τI=2τ2022.11 | 98.23 | — | — | |
| Trans-MCL (EXP)P(s=1)=0.7, Base loss=EXP, Backbone=CNN2026.05 | 98.12 | — | — | |
| Trans-MCL (EXP)P(s=1)=0.3, Base loss=EXP, Backbone=CNN2026.05 | 98.1 | — | — | |
| SNN (temporal coding)Network=784-500-10, Coding=SNN (temporal coding), BP scheme=timing-based, Spike type=single, Leakage=non-leaky, τV=∞, τI=∞2022.11 | 97.99 | — | — | |
| timing-based backpropagation for SNNsNetwork=784-400-10, Coding=SNN (temporal coding), BP scheme=timing-based, Spike type=single, Leakage=leaky, τV=τ, τI=2τ2022.11 | 97.99 | — | — | |
| LILogicNet-SModel=FFN, Gate count=4 K2026.02 | 97.96 | — | — | |
| SNN (temporal coding)Network=784-340-10, Coding=SNN (temporal coding), BP scheme=timing-based, Spike type=single, Leakage=leaky, τV=τ, τI=τ2022.11 | 97.96 | — | — | |
| timing-based backpropagation for SNNsNetwork=784-400-10, Coding=SNN (temporal coding), BP scheme=timing-based, Spike type=single, Leakage=non-leaky*, τV=τ, τI=2τ2022.11 | 97.96 | — | — | |
| FLCNoise Type=Sparse, Noise Level=Low, Bias Correction=false2020.11 | 97.76 | — | — | |
| DiffLogic Net (small)Model=FFN, Gate count=48 K2026.02 | 97.69 | — | — | |
| SNN (temporal coding)Network=784-800-10, Coding=SNN (temporal coding), BP scheme=timing-based, Spike type=single, Leakage=non-leaky, τV=∞, τI=τ2022.11 | 97.55 | — | — | |
| SNN (temporal coding)Network=256-246-10, Coding=SNN (temporal coding), BP scheme=surrogate-gradient, Spike type=multi, Leakage=leaky, τV=τ, τI=τ2022.11 | 97.5 | — | — | |
| BLCNoise Type=Sparse, Noise Level=Low, Bias Correction=false2020.11 | 97.37 | — | — | |
| DRTP# Layers=22025.01 | 97.32 | — | — | |
| PLNoise Type=Sparse, Noise Level=Low, Bias Correction=false2020.11 | 97.21 | — | — | |
| SNN (temporal coding)Network=784-350-10, Coding=SNN (temporal coding), BP scheme=timing-based, Spike type=single, Leakage=leaky, τV=τ, τI=2τ2022.11 | 97.2 | — | — | |
| GATHidden dimension=64, Depth=3, Batch size=16, Dropout=0.1, Global mean pooling=true, Epochs=200, Optimizer=Adam, Learning rate=3 × 10−4, Number of random seeds=52026.05 | 95.94 | — | — | |
| BitLogicModel=CNN, Gate count=253.4 K2026.02 | 95.72 | — | — | |
| ChebConvHidden dimension=64, Depth=3, Batch size=16, Dropout=0.1, Global mean pooling=true, Epochs=200, Optimizer=Adam, Learning rate=3 × 10−4, Number of random seeds=52026.05 | 95.72 | — | — | |
| Trans-MCL (MAE)P(s=1)=0.3, Base loss=MAE, Backbone=CNN2026.05 | 95.49 | — | — | |
| Trans-MCL (MAE)P(s=1)=0.7, Base loss=MAE, Backbone=CNN2026.05 | 95.47 | — | — | |
| CENoise Type=Sparse, Noise Level=Low, Bias Correction=false2020.11 | 95.23 | — | — | |
| Trans-PLL-AvgP(s=1)=0.7, Base loss=MAE, Backbone=CNN2026.05 | 94.85 | — | — | |
| Trans-PLL-AvgP(s=1)=0.5, Base loss=MAE, Backbone=CNN2026.05 | 94.78 | — | — | |
| Trans-PLL-AvgP(s=1)=0.3, Base loss=MAE, Backbone=CNN2026.05 | 93.8 | — | — | |
| GCNHidden dimension=64, Depth=3, Batch size=16, Dropout=0.1, Global mean pooling=true, Epochs=200, Optimizer=Adam, Learning rate=3 × 10−4, Number of random seeds=52026.05 | 92.09 | — | — | |
| PearsonNoise Type=Sparse, Noise Level=High, Bias Correction=false2020.11 | 58.63 | — | — | |
| ADGMData augmentation=false2017.05 | — | — | 962 | |
| CAEN=1002015.06 | — | 13.47 | — | |
| CAEN=6002015.06 | — | 6.3 | — | |
| CAEN=10002015.06 | — | 4.77 | — | |
| CAEN=30002015.06 | — | 3.22 | — | |
| CatGANData augmentation=false2017.05 | — | — | 19,110 | |
| CDBN2014.04 | — | 0.82 | — | |
| Conv. GPInput Dimension=25-dim2019.05 | — | 2.1 | — | |
| Conv. Maxout + Dropout2014.04 | — | 0.45 | — | |
| ConvnetN=1002015.06 | — | 22.98 | — | |
| ConvnetN=6002015.06 | — | 7.86 | — | |
| ConvnetN=10002015.06 | — | 6.45 | — | |
| ConvnetN=30002015.06 | — | 3.35 | — | |
| ConvNet2014.04 | — | 0.53 | — | |
| FMDiscriminator architecture=small, Data augmentation=false2017.05 | — | — | 936.5 | |
| GCGPInput Dimension=24-dim2019.05 | — | 1.7 | — | |
| HSC2014.04 | — | 0.77 | — | |
| K-NN-IDM2014.04 | — | 0.54 | — | |
| K-NN-SCM2014.04 | — | 0.63 | — | |
| Ladder networkData augmentation=false2017.05 | — | — | 10,637 | |
| LDANet-1Filter size (k1, k2)=7, 72014.04 | — | 0.98 | — | |
| LDANet-2Filter size (k1, k2)=7, 72014.04 | — | 0.62 | — | |
| MTCN=1002015.06 | — | 12.03 | — | |
| MTCN=6002015.06 | — | 5.13 | — | |
| MTCN=10002015.06 | — | 3.64 | — | |
| MTCN=30002015.06 | — | 2.57 | — | |
| PCANet-1Filter size (k1, k2)=7, 72014.04 | — | 0.94 | — | |
| PCANet-1 (k1 = 13)Filter size (k1, k2)=13, 72014.04 | — | 0.62 | — | |
| PCANet-2Filter size (k1, k2)=7, 72014.04 | — | 0.66 | — | |
| PL-DAEN=1002015.06 | — | 10.49 | — | |
| PL-DAEN=6002015.06 | — | 5.03 | — |