Image Classification on MNIST (Accuracy and Sparsity)
98.2AccuracySparse VD
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Sparse VDDescription=Variational dropout sparsification, Architecture=512-114-72, Weights Trained=Yes2026.03 | 98.2 | 97.8 | |
| BC-GNJDescription=Bayesian compression with GNJ prior, Architecture=278-98-13, Weights Trained=Yes2026.03 | 98.2 | 89.2 | |
| BC-GHSDescription=Bayesian compression with GHS prior, Architecture=311-86-14, Weights Trained=Yes2026.03 | 98.2 | — | |
| BaselineDescription=Standard training with ℓ2 loss, Architecture=784-300-100, Weights Trained=Yes2026.03 | 96.95 | 0 | |
| CRBGDescription=Continuously relaxed Bernoulli gates, Architecture=784-300-100, Weights Trained=No2026.03 | 96 | 45 | |
| Edge-popupDescription=Score-based selection + ϵ-perturbation, Architecture=500-500-500-500, Weights Trained=No2026.03 | 85 | 50 |