Image Classification on MNIST (train)
100Train AccuracyRandom Forest
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Random Forest#E=100, #S=292k2021.10 | 100 | — | |
| CNNPre-trained=ImageNet, Runs=52026.03 | 100 | — | |
| KMEx (K/C=5)Pre-trained=ImageNet, Runs=52026.03 | 100 | — | |
| ProtoPNetPre-trained=ImageNet, Runs=52026.03 | 100 | — | |
| B4Backbone=ResNet34, Pre-trained=ImageNet, Runs=52026.03 | 100 | — | |
| AdamArchitecture=1 x 128, Steps (k)=1.6 ± 0.5, Time (s)=4.7 ± 1.4, Time/Step (ms)=2.95 ± 0.012025.06 | 100 | — | |
| AdamArchitecture=4 x 128 + Skip, Steps (k)=1.4 ± 0.5, Time (s)=4.3 ± 1.6, Time/Step (ms)=3.08 ± 0.012025.06 | 100 | — | |
| FRE-RNN (tanh, Adam)Architecture=2HL, Epoch / Batch size - T/K=50/500-10/10, Wall Clock Time=0:01:162025.08 | 100 | — | |
| BP (tanh, Adam)Architecture=2HL, Epoch / Batch size - T/K=50/500-1/1, Wall Clock Time=0:00:182025.08 | 100 | — | |
| FRE-RNN (tanh, Adam)Architecture=3HL, Epoch / Batch size - T/K=50/500-18/10, Wall Clock Time=0:02:112025.08 | 100 | — | |
| BP (tanh, Adam)Architecture=3HL, Epoch / Batch size - T/K=50/500-1/1, Wall Clock Time=0:00:242025.08 | 100 | — | |
| FRE-RNN (tanh)Architecture=3HL, Epoch / Batch size - T/K=100/20-18/10, Wall Clock Time=1:01:542025.08 | 99.98 | — | |
| AdamArchitecture=4 x 128, Steps (k)=1.8 ± 0.7, Time (s)=5.7 ± 2.3, Time/Step (ms)=3.09 ± 0.052025.06 | 99.9 | — | |
| P-EP (sigmoid-s)Architecture=3HL, Epoch / Batch size - T/K=100/20-180/20, Wall Clock Time=8:27:-2025.08 | 99.9 | — | |
| P-EP (sigmoid-s)Architecture=2HL, Epoch / Batch size - T/K=50/20-100/20, Wall Clock Time=1:56:-2025.08 | 99.86 | — | |
| FRE-RNN (hard-sigmoid)Architecture=Conv, Epoch / Batch size - T/K=40/128-20/10, Wall Clock Time=0:12:282025.08 | 99.78 | — | |
| MLPHyperparameters=h = 64, Number of Parameters=148,586, Architecture=convolutional frontend (2304–64–10)2026.06 | 99.62 | — | |
| ECCEncoding=one-hot2017.04 | 99.53 | — | |
| JSD-LBepsilon=0.12025.10 | 99.5 | — | |
| FAArchitecture=1 x 128, Steps (k)=8.0 ± 1.1, Time (s)=23.3 ± 3.1, Time/Step (ms)=2.92 ± 0.022025.06 | 99.5 | — | |
| P-EP (hard-sigmoid)Architecture=Conv, Epoch / Batch size - T/K=40/20-200/10, Wall Clock Time=8:58:-2025.08 | 99.46 | — | |
| BP (hard-sigmoid)Architecture=Conv, Epoch / Batch size - T/K=40/128-1/1, Wall Clock Time=0:01:012025.08 | 99.43 | — | |
| ECCInput type=sparse2017.04 | 99.36 | — | |
| LPNumber of Parameters=23,466, Architecture=convolutional frontend (2304–64–10)2026.06 | 99.2 | — | |
| ECCInput type=regular2017.04 | 99.12 | — | |
| MLPHyperparameters=h = 20, Number of Parameters=46,726, Architecture=convolutional frontend (2304–64–10)2026.06 | 99.12 | — | |
| AdamArchitecture=4 x 16 + Skip2025.06 | 99.1 | — | |
| TnT-bagging#E=100, #S=111k2021.10 | 99.09 | — | |
| NAS2025.02 | 99.07 | — | |
| AdamArchitecture=4 x 162025.06 | 99 | — | |
| Baseline network2025.02 | 98.98 | — | |
| FRE-RNN-DLR (tanh)Architecture=3HL, Epoch / Batch size - T/K=100/20-18/10, Wall Clock Time=1:01:142025.08 | 98.93 | — | |
| Proposed (I)Type=I2025.02 | 98.92 | — | |
| Net2DeeperNet (II)Type=II2025.02 | 98.92 | — | |
| Proposed (II)Type=II2025.02 | 98.9 | — | |
| Random layer insertion (I)Type=I2025.02 | 98.9 | — | |
| Forward Thinking2025.02 | 98.9 | — | |
| B234Pre-trained=ImageNet, Runs=52026.03 | 98.9 | — | |
| HalfNetHyperparameters=k = 8, Number of Parameters=21,930, Architecture=convolutional frontend (2304–64–10)2026.06 | 98.9 | — | |
| SGDArchitecture=4 x 128, Steps (k)=32.4 ± 2.3, Time (s)=96.2 ± 7.2, Time/Step (ms)=2.97 ± 0.032025.06 | 98.8 | — | |
| DBPTTModel variant=DBPTT2026.05 | 98.76 | — | |
| MLPNumber of Parameters=55,050, Architecture=784–64–64–102026.06 | 98.71 | — | |
| TnT-bagging#E=20, #S=19.2k2021.10 | 98.64 | — | |
| EPModel variant=EP2026.05 | 98.63 | — | |
| DEPModel variant=DEP2026.05 | 98.6 | — | |
| MLPHyperparameters=h = 10, Number of Parameters=23,576, Architecture=convolutional frontend (2304–64–10)2026.06 | 98.6 | — | |
| HalfNetHyperparameters=k = 4, Number of Parameters=12,714, Architecture=convolutional frontend (2304–64–10)2026.06 | 98.46 | — | |
| MLPHyperpars=h = 64, Params=50,8902026.06 | 98.36 | — | |
| TnT-bagging#E=10, #S=9.6k2021.10 | 98.28 | — | |
| HalfNetHyperparameters=k = 48, Number of Parameters=42,458, Architecture=784–64–64–102026.06 | 98.15 | — | |
| SGDArchitecture=4 x 128 + Skip, Steps (k)=28.2 ± 7.2, Time (s)=84.9 ± 21.9, Time/Step (ms)=3.01 ± 0.012025.06 | 98.1 | — | |
| TnT-AdaBoost#E=20, #S=2.9k2021.10 | 98.03 | — | |
| GEGONumber of Neurons=[378, 191, 220, 106], Dropout Rate=[0, 0.1, 0, 0], Batch Size=224, Learning Rate=0.012026.01 | 97.95 | — | |
| Random Forest#E=20, #S=19.2k2021.10 | 97.9 | — | |
| GEONumber of Neurons=[512, 135, 504, 220], Dropout Rate=[0.5, 0.5, 0, 0], Batch Size=32, Learning Rate=0.012026.01 | 97.85 | — | |
| GANumber of Neurons=[356, 388, 413, 80], Dropout Rate=[0.1, 0, 0, 0], Batch Size=224, Learning Rate=0.012026.01 | 97.76 | — | |
| AdaBoost#E=20, #S=2.9k2021.10 | 97.7 | — | |
| HalfNetHyperparameters=k = 2, Number of Parameters=8,106, Architecture=convolutional frontend (2304–64–10)2026.06 | 97.7 | — | |
| HalfNetHyperparameters=k = 32, Number of Parameters=28,890, Architecture=784–64–64–102026.06 | 97.67 | — | |
| SGDArchitecture=1 x 128, Steps (k)=58.6 ± 7.2, Time (s)=171.4 ± 21.4, Time/Step (ms)=2.92 ± 0.012025.06 | 97.6 | — | |
| TnT-bagging#E=5, #S=4.8k2021.10 | 97.46 | — | |
| Random Forest#E=10, #S=9.6k2021.10 | 97.44 | — | |
| SGDArchitecture=4 x 16 + Skip2025.06 | 97.1 | — | |
| HalfNetHyperparameters=k = 1, Number of Parameters=5,802, Architecture=convolutional frontend (2304–64–10)2026.06 | 96.98 | — | |
| DRArchitecture=1 x 128, Steps (k)=16.6 ± 3.3, Time (s)=3.7 ± 0.7, Time/Step (ms)=0.22 ± 0.002025.06 | 96.9 | — | |
| Random Forest#E=5, #S=4.8k2021.10 | 96.55 | — | |
| HalfNetHyperpars=k = 16, Params=14,1062026.06 | 96.39 | — | |
| HalfNetHyperparameters=k = 16, Number of Parameters=15,322, Architecture=784–64–64–102026.06 | 96.35 | — | |
| SGDArchitecture=4 x 162025.06 | 96.3 | — | |
| FF (Adam)Architecture=1 x 128, Steps (k)=82.4 ± 11.8, Time (s)=256.2 ± 35.2, Time/Step (ms)=3.12 ± 0.042025.06 | 96.2 | — | |
| FAArchitecture=4 x 128, Steps (k)=20.0 ± 6.2, Time (s)=60.2 ± 18.3, Time/Step (ms)=3.02 ± 0.032025.06 | 96.1 | — | |
| DRArchitecture=4 x 128 + Skip, Steps (k)=21.8 ± 6.0, Time (s)=9.3 ± 2.6, Time/Step (ms)=0.43 ± 0.002025.06 | 96.1 | — | |
| JSD-LBepsilon=0.22025.10 | 96 | — | |
| MLPHyperpars=h = 20, Params=15,9102026.06 | 95.63 | — | |
| HalfNetHyperpars=k = 16, Params=14,106, Architecture=784-64-10, Weights=binary half-layer weights2026.06 | 95.38 | — | |
| TnT-AdaBoost#E=10, #S=1.4k2021.10 | 95.09 | — | |
| HalfNetHyperpars=k = 8, Params=7,8342026.06 | 95.06 | — | |
| HalfNetHyperparameters=k = 8, Number of Parameters=8,538, Architecture=784–64–64–102026.06 | 94.66 | — | |
| DisenIBepsilon=0.12025.10 | 94.3 | — | |
| AdaBoost#E=10, #S=1.4k2021.10 | 94.28 | — | |
| FF (Adam)Architecture=4 x 128, Steps (k)=66.0 ± 9.3, Time (s)=220.3 ± 32.1, Time/Step (ms)=3.35 ± 0.062025.06 | 94.2 | — | |
| Adaptive CFF (Learnable)collaboration_parameters=learnable, initial_gamma_l=1, learning_rate_gamma=0.012025.12 | 94.1 | — | |
| HalfNetHyperpars=k = 8, Params=7,834, Architecture=784-64-10, Weights=binary half-layer weights2026.06 | 93.96 | — | |
| MLPHyperpars=h = 10, Params=7,9602026.06 | 93.36 | — | |
| HalfNetHyperpars=k = 4, Params=4,6982026.06 | 93.11 | — | |
| randomNumber of Parameters=1,066, Architecture=convolutional frontend (2304–64–10)2026.06 | 93.08 | — | |
| LPParams=7,8502026.06 | 92.84 | — | |
| HalfNetHyperparameters=k = 4, Number of Parameters=5,146, Architecture=784–64–64–102026.06 | 92.74 | — | |
| Collaborative FF (Fixed)collaboration_parameters=fixed, gamma_l=12025.12 | 92.5 | — | |
| CPArchitecture=1 x 128, Steps (k)=41.6 ± 11.4, Time (s)=8.6 ± 2.3, Time/Step (ms)=0.21 ± 0.002025.06 | 92.5 | — | |
| Original FFcollaboration=none2025.12 | 91.8 | — | |
| JSD-LBepsilon=0.32025.10 | 91.4 | — | |
| APArchitecture=1 x 128, Steps (k)=25.8 ± 5.4, Time (s)=2.3 ± 0.5, Time/Step (ms)=0.09 ± 0.002025.06 | 91.4 | — | |
| HalfNetHyperpars=k = 2, Params=3,1302026.06 | 91.33 | — | |
| CPArchitecture=4 x 128 + Skip, Steps (k)=34.2 ± 3.5, Time (s)=18.7 ± 1.9, Time/Step (ms)=0.55 ± 0.002025.06 | 91.1 | — | |
| HalfNetHyperpars=k = 1, Params=2,3462026.06 | 90.38 | — | |
| APArchitecture=4 x 128 + Skip, Steps (k)=32.6 ± 9.3, Time (s)=6.9 ± 2.0, Time/Step (ms)=0.21 ± 0.002025.06 | 90.3 | — | |
| TnT-AdaBoost#E=5, #S=6402021.10 | 90.26 | — | |
| HalfNetHyperparameters=k = 2, Number of Parameters=3,450, Architecture=784–64–64–102026.06 | 89.93 | — | |
| AdaBoost#E=5, #S=6402021.10 | 89.75 | — |