Image Classification on ImageNet-1K (train val) with Internal Layer Metrics
19.3Conv1 Activation ValueImageNet-labels
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| ImageNet-labelsProtocol=Linear logistic regression (Linear probing), Weights=Supervised labels, Backbone=AlexNet, Feature Map Dim=~90002016.11 | 19.3 | 36.3 | 44.2 | 48.3 | 50.5 | |
| Noroozi & FavaroProtocol=Linear logistic regression (Linear probing), Weights=Frozen, Backbone=AlexNet variant, conv1 stride=2, Feature Map Dim=~90002016.11 | 19.2 | 30.1 | 34.7 | 33.9 | 28.3 | |
| Donahue et al.Protocol=Linear logistic regression (Linear probing), Weights=Frozen, Backbone=AlexNet variant, Feature Map Dim=~90002016.11 | 17.7 | 24.5 | 31 | 29.9 | 28 | |
| Split-Brain Auto (cl,cl)Protocol=Linear logistic regression (Linear probing), Weights=Frozen, Backbone=AlexNet variant, Feature Map Dim=~9000, Loss type=Classification (cl,cl)2016.11 | 17.7 | 29.3 | 35.4 | 35.2 | 32.8 | |
| Krähenbühl et al.Protocol=Linear logistic regression (Linear probing), Weights=Frozen, Backbone=AlexNet variant, Feature Map Dim=~90002016.11 | 17.5 | 23 | 24.5 | 23.2 | 20.6 | |
| Split-Brain Auto (reg,reg)Protocol=Linear logistic regression (Linear probing), Weights=Frozen, Backbone=AlexNet variant, Feature Map Dim=~9000, Loss type=Regression (reg,reg)2016.11 | 17.4 | 27.9 | 33.6 | 34.2 | 32.3 | |
| Doersch et al.Protocol=Linear logistic regression (Linear probing), Weights=Frozen, Backbone=AlexNet variant, Feature Map Dim=~90002016.11 | 16.2 | 23.3 | 30.2 | 31.7 | 29.6 | |
| (L,ab,Lab)->(ab,L,Lab)Protocol=Linear logistic regression (Linear probing), Weights=Frozen, Backbone=AlexNet variant, Feature Map Dim=~90002016.11 | 15.4 | 22.9 | 24 | 22 | 18.9 | |
| (L,ab)->(ab,L)Protocol=Linear logistic regression (Linear probing), Weights=Frozen, Backbone=AlexNet variant, Feature Map Dim=~90002016.11 | 15.1 | 22.6 | 24.4 | 23.2 | 21.1 | |
| Pathak et al.Protocol=Linear logistic regression (Linear probing), Weights=Frozen, Backbone=AlexNet variant, Feature Map Dim=~90002016.11 | 14.1 | 20.7 | 21 | 19.8 | 15.5 | |
| Zhang et al.Protocol=Linear logistic regression (Linear probing), Weights=Frozen, Backbone=AlexNet variant, Feature Map Dim=~90002016.11 | 13.1 | 24.8 | 31 | 32.6 | 31.8 | |
| Lab->LabProtocol=Linear logistic regression (Linear probing), Weights=Frozen, Backbone=AlexNet variant, Feature Map Dim=~9000, Loss=L2 Regression2016.11 | 12.9 | 20.1 | 18.5 | 15.1 | 11.5 | |
| L->ab(cl)Protocol=Linear logistic regression (Linear probing), Weights=Frozen, Backbone=AlexNet variant, Feature Map Dim=~9000, Task=Classification-based colorization2016.11 | 12.5 | 25.4 | 32.4 | 33.1 | 32 | |
| L->ab(reg)Protocol=Linear logistic regression (Linear probing), Weights=Frozen, Backbone=AlexNet variant, Feature Map Dim=~9000, Task=Regression-based colorization2016.11 | 12.3 | 23.5 | 29.6 | 31.1 | 30.1 | |
| Lab(drop50)->LabProtocol=Linear logistic regression (Linear probing), Weights=Frozen, Backbone=AlexNet variant, Feature Map Dim=~9000, Denoising=50% Dropout2016.11 | 12.1 | 20.4 | 19.7 | 16.1 | 12.3 | |
| Ensembled L->abProtocol=Linear logistic regression (Linear probing), Weights=Frozen, Backbone=AlexNet variant, Feature Map Dim=~90002016.11 | 11.7 | 23.7 | 30.9 | 32.2 | 31.3 | |
| GaussianProtocol=Linear logistic regression (Linear probing), Weights=Random initialization, Backbone=AlexNet variant, Feature Map Dim=~90002016.11 | 11.6 | 17.1 | 16.9 | 16.3 | 14.1 | |
| ab->L(cl)Protocol=Linear logistic regression (Linear probing), Weights=Frozen, Backbone=AlexNet variant, Feature Map Dim=~9000, Task=Classification-based grayscale prediction2016.11 | 11.6 | 19.2 | 22.6 | 21.7 | 19.2 | |
| ab->L(reg)Protocol=Linear logistic regression (Linear probing), Weights=Frozen, Backbone=AlexNet variant, Feature Map Dim=~9000, Task=Regression-based grayscale prediction2016.11 | 11.5 | 19.4 | 23.5 | 23.9 | 21.7 |