Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Masksembles for Uncertainty Estimation
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace randomly resampled dropout masks with a fixed pool of structured, controllably overlapping masks | MC-D on Wide-ResNet-16-4 · CIFAR10 vs SVHN OOD | OOD ROC: 0.92 -> 0.94 | Helped |
| Replace randomly resampled dropout masks with a fixed pool of structured, controllably overlapping masks | MC-D on Wide-ResNet-16-4 · CIFAR10 vs SVHN OOD | OOD PR: 0.95 -> 0.96 | Helped |
| Use Masksembles instead of MC-Dropout | Wide-ResNet-16-4 · CIFAR10 test | accuracy: 0.90 -> 0.92 | Helped |
| Lower the mask overlap by varying scale while holding total model size fixed (ablation) | Masksembles on CIFAR10 · CIFAR10-C corruptions | ECE | Helped |
| Increase mask overlap scale from IoU 0.7 to IoU 0.3 (ablation) | Masksembles on ResNet-50 · ImageNet test | ECE: 0.03 -> 0.02 | Helped |
| Decrease mask overlap scale from IoU 0.3 to IoU 0.2 (ablation) | Masksembles on ResNet-50 · ImageNet test | ECE: 0.02 -> 0.02 | No clear effect |
| Decrease mask overlap scale from IoU 0.7 to IoU 0.3 (ablation) | Masksembles on ResNet-50 · ImageNet test | accuracy: 0.72 -> 0.71 | Hurt |
| Decrease mask overlap scale from IoU 0.3 to IoU 0.2 (ablation) | Masksembles on ResNet-50 · ImageNet test | accuracy: 0.71 -> 0.70 | Hurt |
Closest papers by meaning
- Neural Ensemble Search for Uncertainty Estimation and Dataset Shift (NeurIPS 2021 · similarity 0.86)
- Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles (NeurIPS 2017 · similarity 0.86)
- Packed Ensembles for efficient uncertainty estimation (ICLR 2023 · similarity 0.86)
- Contextual Dropout: An Efficient Sample-Dependent Dropout Module (ICLR 2021 · similarity 0.85)
- Pitfalls of In-Domain Uncertainty Estimation and Ensembling in Deep Learning (ICLR 2020 · similarity 0.85)
- SURE: SUrvey REcipes for building reliable and robust deep networks (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.