Image Classification on ImageNet (test val)
78.6AccuracyBYOL
Evaluation results
| Method | Links | ||
|---|---|---|---|
| BYOLPre-training category=Discriminative Pretraining, Evaluation protocol=Linear-probe, Features=8192, Params=375M2021.10 | 78.6 | — | |
| BaselineW-bit=32, A-bit=322020.01 | 77.72 | 97.49 | |
| ZEROQNo D=true, No FT=true, W-bit=8, A-bit=82020.01 | 77.67 | 24.37 | |
| ZEROQNo D=true, No FT=true, W-bit=MP, A-bit=62020.01 | 77.43 | 18.27 | |
| PACTNo D=false, No FT=false, W-bit=4, A-bit=42020.01 | 76.5 | 12.19 | |
| SimCLRPre-training category=Discriminative Pretraining, Evaluation protocol=Linear-probe, Features=8192, Params=375M2021.10 | 76.5 | — | |
| ZEROQNo D=true, No FT=true, W-bit=MP, A-bit=8, Percentile Quantization=true2020.01 | 76.08 | 12.17 | |
| ZEROQNo D=true, No FT=true, W-bit=MP, A-bit=82020.01 | 75.8 | 12.17 | |
| HAWQ-V2Backbone=ResNet-50, ACR=8, WCR=12.242020.07 | 75.7 | — | |
| HMQBackbone=ResNet-50, ACR=8, WCR=13.1, Target weight compression rate (Rw)=132020.07 | 75.45 | — | |
| HAWQBackbone=ResNet-50, ACR=8, WCR=12.282020.07 | 75.3 | — | |
| SwAVPre-training category=Discriminative Pretraining, Evaluation protocol=Linear-probe, Features=2048, Params=93M2021.10 | 75.3 | — | |
| DINOPre-training category=Discriminative Pretraining, Evaluation protocol=Linear-probe, Features=2048, Params=85M2021.10 | 75.3 | — | |
| OMSENo D=false, No FT=true, W-bit=4, A-bit=322020.01 | 74.98 | 12.28 | |
| OCSNo D=false, No FT=true, W-bit=6, A-bit=62020.01 | 74.8 | 18.46 | |
| Spiking ResNet-34 [33]Training Method=ANN2SNN, T=2562021.02 | 74.61 | — | |
| FPBackbone=ResNet-34, Bit-width=32W32A2020.09 | 73.3 | — | |
| VIM-LargePre-training category=Generative Pretraining, Evaluation protocol=Linear-probe, Notes=trained without dropout in transformers, #Tokens=32 x 32, Features=2048, Params=1697M2021.10 | 73.2 | — | |
| Spiking ResNet-50 [17]Training Method=ANN2SNN, T=3502021.02 | 72.75 | — | |
| iGPT-XLPre-training category=Generative Pretraining, Evaluation protocol=Linear-probe, Training Mode=extra data, feature ensemble, #Tokens=64 x 64, Features=5 x 3072, Params=6801M2021.10 | 72 | — | |
| QuantNetBackbone=ResNet-34, Bit-width=1W32A2020.09 | 71.97 | — | |
| CPC v2Pre-training category=Discriminative Pretraining, Evaluation protocol=Linear-probe, Features=4096, Params=303M2021.10 | 71.5 | — | |
| QuantNet(D)Backbone=ResNet-34, Bit-width=1W32A2020.09 | 71.35 | — | |
| Meta-QuantBackbone=ResNet-34, Bit-width=1W32A2020.09 | 70.84 | — | |
| ProxQuantBackbone=ResNet-34, Bit-width=1W32A2020.09 | 70.42 | — | |
| OMSENo D=true, No FT=true, W-bit=4, A-bit=322020.01 | 70.06 | 12.28 | |
| Spiking ResNet-34 [12]Training Method=ANN2SNN, T=40962021.02 | 69.89 | — | |
| SEW ResNet-152Training Method=Spike-based BP, T=42021.02 | 69.26 | — | |
| SEW ResNet-101Training Method=Spike-based BP, T=42021.02 | 68.76 | — | |
| iGPT-XLPre-training category=Generative Pretraining, Evaluation protocol=Linear-probe, Training Mode=extra data, #Tokens=64 x 64, Features=3072, Params=6801M2021.10 | 68.7 | — | |
| MoCoPre-training category=Discriminative Pretraining, Evaluation protocol=Linear-probe, Features=8192, Params=375M2021.10 | 68.6 | — | |
| AMDIMPre-training category=Discriminative Pretraining, Evaluation protocol=Linear-probe, Features=8192, Params=626M2021.10 | 68.1 | — | |
| SEW ResNet-50Training Method=Spike-based BP, T=42021.02 | 67.78 | — | |
| Spiking ResNet-34 (large) with td-BNTraining Method=Spike-based BP, T=6, Notes=Uses 4x convolution kernels2021.02 | 67.05 | — | |
| SEW ResNet-34Training Method=Spike-based BP, T=42021.02 | 67.04 | — | |
| Spiking ResNet-34 [49]Training Method=ANN2SNN, T=20002021.02 | 65.47 | — | |
| iGPT-LPre-training category=Generative Pretraining, Evaluation protocol=Linear-probe, #Tokens=48 x 48, Features=1536, Params=1362M2021.10 | 65.2 | — | |
| VIM-BasePre-training category=Generative Pretraining, Evaluation protocol=Linear-probe, #Tokens=32 x 32, Features=1024, Params=650M2021.10 | 65.1 | — | |
| Spiking ResNet-50 with td-BNTraining Method=Spike-based BP, T=62021.02 | 64.88 | — | |
| VIM-Base + DALL-E dVAE quantizerPre-training category=Generative Pretraining, Evaluation protocol=Linear-probe, #Tokens=32 x 32, Features=1024, Params=650M, Quantizer=DALL-E dVAE2021.10 | 63.8 | — | |
| Spiking ResNet-34 with td-BNTraining Method=Spike-based BP, T=62021.02 | 63.72 | — | |
| VIM-Base + CNN-VQGAN quantizerPre-training category=Generative Pretraining, Evaluation protocol=Linear-probe, #Tokens=32 x 32, Features=1024, Params=650M, Quantizer=CNN-VQGAN2021.10 | 61.8 | — | |
| Spiking ResNet-34 [43]Training Method=ANN2SNN and Spike-based BP, T=2502021.02 | 61.48 | — | |
| BigBiGANPre-training category=Generative Pretraining, Evaluation protocol=Linear-probe, Features=16384, Params=344M2021.10 | 61.3 | — | |
| iGPT-LPre-training category=Generative Pretraining, Evaluation protocol=Linear-probe, #Tokens=32 x 32, Features=1536, Params=1362M2021.10 | 60.3 | — | |
| 0.50 MobileNet-160Million Mult-Adds=76, Million Parameters=1.32, Width Multiplier=0.5, Resolution=160x1602017.04 | 60.2 | — | |
| SqueezenetMillion Mult-Adds=1700, Million Parameters=1.252017.04 | 57.5 | — | |
| AlexNetMillion Mult-Adds=720, Million Parameters=602017.04 | 57.2 | — | |
| BigBiGANPre-training category=Generative Pretraining, Evaluation protocol=Linear-probe, Features=4096, Params=86M2021.10 | 56.6 | — | |
| RotationPre-training category=Discriminative Pretraining, Evaluation protocol=Linear-probe, Features=8192, Params=86M2021.10 | 55.4 | — | |
| FPBackbone=AlexNet, Bit-width=32W32A2020.09 | 55.07 | — | |
| QuantNetBackbone=AlexNet, Bit-width=1W32A2020.09 | 54.06 | — | |
| QuantNet(D)Backbone=AlexNet, Bit-width=1W32A2020.09 | 53.59 | — | |
| Self-Binar.Backbone=AlexNet, Bit-width=1W32A2020.09 | 52.89 | — | |
| RelativePositionPre-training category=Discriminative Pretraining, Evaluation protocol=Linear-probe, Features=4096, Params=94M2021.10 | 51.4 | — | |
| Self-Binar.(D)Backbone=AlexNet, Bit-width=1W32A2020.09 | 50.51 | — | |
| JigsawPre-training category=Discriminative Pretraining, Evaluation protocol=Linear-probe, Features=4096, Params=94M2021.10 | 44.6 | — | |
| BiGANPre-training category=Generative Pretraining, Evaluation protocol=Linear-probe, Features=512, Params=138M2021.10 | 31 | — |