Image Classification on MNIST
99.87AccuracySOTA
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| SOTA2024.06 | 99.87 | — | — | — | |
| Conditional Channel Gated NetworksEvaluation Protocol=Task-IL, Number of runs=52020.03 | 99.8 | — | — | — | |
| Multiple modelsBackbone=ViT-L/142022.08 | 99.8 | — | — | — | |
| ResNet26-64Parameter Size=0.89MP2024.10 | 99.73 | — | — | — | |
| HATEvaluation Protocol=Task-IL, Number of runs=52020.03 | 99.7 | — | — | — | |
| Multiple modelsBackbone=ViT-B/162022.08 | 99.7 | — | — | — | |
| Joint patchingBackbone=ViT-L/142022.08 | 99.7 | — | — | — | |
| R-ExplaiNet26-64Parameter Size=0.89MP2024.10 | 99.7 | — | — | — | |
| ConvPointInput type=Point-based, Data variant=Gray levels, Neighborhood size (K)=162019.04 | 99.62 | — | — | — | |
| Multiple modelsBackbone=ViT-B/322022.08 | 99.6 | — | — | — | |
| PointCNNInput type=Point-based2019.04 | 99.54 | — | — | — | |
| NiNInput type=Image-based2019.04 | 99.53 | — | — | — | |
| AT + IGRAttack Protocol=Natural2022.05 | 99.51 | — | — | — | |
| MART + IGRAttack Protocol=Natural2022.05 | 99.51 | — | — | — | |
| PointNet++Input type=Point-based2019.04 | 99.49 | — | — | — | |
| ConvPointInput type=Point-based, Data variant=Black points, Neighborhood size (K)=162019.04 | 99.49 | — | — | — | |
| AT + IGRAttack Protocol=FGSM2022.05 | 99.45 | — | — | — | |
| ATAttack Protocol=Natural2022.05 | 99.43 | — | — | — | |
| STBPArchitecture=Conv, 5 layers2026.03 | 99.42 | — | — | — | |
| TRADESAttack Protocol=Natural2022.05 | 99.4 | — | — | — | |
| TRADES + IGRAttack Protocol=Natural2022.05 | 99.4 | — | — | — | |
| TRADES + IGRAttack Protocol=FGSM2022.05 | 99.4 | — | — | — | |
| Zhang et al. (2021)Spike code=single, Architecture=16C5-P2-32C5-P2-800-128-10, Neuron model=IF (ReL-PSP)2022.05 | 99.4 | — | — | — | |
| PLMNoise Type=Pairwise, Noise Rate=20%2024.05 | 99.4 | 0.03 | — | — | |
| ATAttack Protocol=FGSM2022.05 | 99.39 | — | — | — | |
| MARTAttack Protocol=Natural2022.05 | 99.39 | — | — | — | |
| MART + IGRAttack Protocol=FGSM2022.05 | 99.39 | — | — | — | |
| PLMNoise Type=Pairwise, Noise Rate=45%2024.05 | 99.39 | 0.05 | — | — | |
| SupCon2023.06 | 99.38 | — | — | — | |
| TRADESAttack Protocol=FGSM2022.05 | 99.36 | — | — | — | |
| AdaNormk=1/102019.11 | 99.35 | — | — | — | |
| MP-FedKDNumber of clients=102026.03 | 99.33 | — | — | — | |
| AT + IGRAttack Protocol=PGD202022.05 | 99.32 | — | — | — | |
| AT + IGRAttack Protocol=CW∞2022.05 | 99.32 | — | — | — | |
| PLMNoise Type=Symmetric, Noise Rate=20%2024.05 | 99.32 | 0.02 | — | — | |
| Wilson-Cowan model for metapopulationBackbone=simple CNN2024.06 | 99.31 | 0.0008 | — | — | |
| Nested StrategyCorruption=15% GN2026.02 | 99.3 | — | — | — | |
| Nested StrategyCorruption=Gaussian Blur2026.02 | 99.3 | — | — | — | |
| MARTAttack Protocol=FGSM2022.05 | 99.29 | — | — | — | |
| MART + IGRAttack Protocol=PGD202022.05 | 99.28 | — | — | — | |
| TRADES + IGRAttack Protocol=PGD202022.05 | 99.26 | — | — | — | |
| ATAttack Protocol=PGD202022.05 | 99.25 | — | — | — | |
| ATAttack Protocol=CW∞2022.05 | 99.24 | — | — | — | |
| TRADES + IGRAttack Protocol=CW∞2022.05 | 99.24 | — | — | — | |
| MART + IGRAttack Protocol=CW∞2022.05 | 99.24 | — | — | — | |
| IndividualBackbone=ViT-B/16, Pre-trained=ImageNet-21k, Number of Merged Models=12024.05 | 99.22 | — | — | — | |
| TRADESAttack Protocol=PGD202022.05 | 99.21 | — | — | — | |
| FACTORCL-SUP2023.06 | 99.21 | — | — | — | |
| LeNetInput type=Image-based2019.04 | 99.2 | — | — | — | |
| Mirsadeghi et al. (2021)Spike code=single, Architecture=40C5-P2-1000-10, Neuron model=IF (PL-PSP)2022.05 | 99.2 | — | — | — | |
| Joint patchingBackbone=ViT-B/322022.08 | 99.2 | — | — | — | |
| Joint patchingBackbone=ViT-B/162022.08 | 99.2 | — | — | — | |
| Parallel StrategyCorruption=15% GN2026.02 | 99.2 | — | — | — | |
| Parallel StrategyCorruption=30% GN2026.02 | 99.2 | — | — | — | |
| Parallel StrategyCorruption=Gaussian Blur2026.02 | 99.2 | — | — | — | |
| TRADESAttack Protocol=CW∞2022.05 | 99.19 | — | — | — | |
| S2-STDP+NCGNeurons per class=5, Feature Extractor=SoftHebb-CNN2024.10 | 99.17 | — | — | — | |
| w/o Norm2019.11 | 99.14 | — | — | — | |
| LayerNorm2019.11 | 99.13 | — | — | — | |
| EuclideanEmbedding Space=Euclidean2021.07 | 99.12 | — | — | — | |
| Class2SimiNoise Type=Symmetric, Noise Rate=20%2024.05 | 99.11 | 0.06 | — | — | |
| VolMinNetNoise Type=Pairwise, Noise Rate=20%2024.05 | 99.11 | 0.08 | — | — | |
| MP-FedKDNumber of clients=202026.03 | 99.11 | — | — | — | |
| VolMinNetNoise Type=Pairwise, Noise Rate=45%2024.05 | 99.1 | 0.08 | — | — | |
| Nested StrategyCorruption=30% GN2026.02 | 99.1 | — | — | — | |
| LayerNorm-simple2019.11 | 99.09 | — | — | — | |
| MARTAttack Protocol=PGD202022.05 | 99.09 | — | — | — | |
| C-HyperbolicEmbedding Space=Clipped Hyperbolic2021.07 | 99.08 | — | — | — | |
| MARTAttack Protocol=CW∞2022.05 | 99.08 | — | — | — | |
| BaselineBackbone=LeNet5, Rel. perturbation ε=0, c.r. (%)=02023.06 | 99.08 | — | — | — | |
| T-RevisionNoise Type=Pairwise, Noise Rate=20%2024.05 | 99.07 | 0.1 | — | — | |
| DLRTBackbone=LeNet5, Rel. perturbation ε=0, c.r. (%)=502023.06 | 99.07 | — | — | — | |
| MOONNumber of clients=102026.03 | 99.02 | — | — | — | |
| CLIPLinear evaluation=true, Model size=Base, Patch size=16*16, Resolution=224*2242023.01 | 99 | — | — | — | |
| FLAVALinear evaluation=true, Model size=Base, Patch size=16*16, Resolution=224*2242023.01 | 99 | — | — | — | |
| DaVinciLinear evaluation=true, Model size=Base, Patch size=16*16, Resolution=224*2242023.01 | 99 | — | — | — | |
| X-FM_baseLinear evaluation=true, Model size=Base, Patch size=16*16, Resolution=224*2242023.01 | 99 | — | — | — | |
| MP-FedKDNumber of clients=502026.03 | 99 | — | — | — | |
| EMR-MERGINGBackbone=ViT-B/16, Pre-trained=ImageNet-21k, Number of Merged Models=302024.05 | 98.99 | — | — | — | |
| Class2SimiNoise Type=Pairwise, Noise Rate=20%2024.05 | 98.97 | 0.12 | — | — | |
| PLMNoise Type=Symmetric, Noise Rate=50%2024.05 | 98.96 | 0.07 | — | — | |
| SSTDP+NCGNeurons per class=5, Feature Extractor=SoftHebb-CNN2024.10 | 98.96 | — | — | — | |
| IDGPBackbone=LeNet-5, Setting=Instance-dependent PLL, Data Augmentation=Yes2022.04 | 98.92 | — | — | — | |
| S2-STDP+NCGNeurons per class=5, Feature Extractor=STDP-CSNN2024.10 | 98.92 | — | — | — | |
| Alternating StrategyCorruption=15% GN2026.02 | 98.9 | — | — | — | |
| DLRTBackbone=LeNet5, Rel. perturbation ε=0, c.r. (%)=802023.06 | 98.88 | — | — | — | |
| DMINoise Type=Symmetric, Noise Rate=20%2024.05 | 98.87 | 0.09 | — | — | |
| T-RevisionNoise Type=Symmetric, Noise Rate=20%2024.05 | 98.86 | 0.04 | — | — | |
| Dual-TNoise Type=Pairwise, Noise Rate=20%2024.05 | 98.86 | 0.09 | — | — | |
| TACArchitecture=Conv, 2 layers, Hardware=M3 Max, K=82026.03 | 98.85 | — | — | — | |
| CAVLBackbone=LeNet-5, Setting=Instance-dependent PLL, Data Augmentation=Yes2022.04 | 98.84 | — | — | — | |
| SimCLR2023.06 | 98.84 | — | — | — | |
| CondLRBackbone=LeNet5, Rel. perturbation ε=0, c.r. (%)=50, τ=02023.06 | 98.84 | — | — | — | |
| Dual-TNoise Type=Symmetric, Noise Rate=20%2024.05 | 98.82 | 0.17 | — | — | |
| VolMinNetNoise Type=Symmetric, Noise Rate=20%2024.05 | 98.82 | 0.12 | — | — | |
| ForwardNoise Type=Pairwise, Noise Rate=20%2024.05 | 98.82 | 0.15 | — | — | |
| DMINoise Type=Pairwise, Noise Rate=20%2024.05 | 98.81 | 0.13 | — | — | |
| S2-STDPNeurons per class=1, Feature Extractor=SoftHebb-CNN2024.10 | 98.81 | — | — | — | |
| CondLRBackbone=LeNet5, Rel. perturbation ε=0, c.r. (%)=80, τ=02023.06 | 98.81 | — | — | — | |
| CondLRBackbone=LeNet5, Rel. perturbation ε=0, c.r. (%)=80, τ=0.12023.06 | 98.77 | — | — | — |