Research map / Architectures and efficiency / CNN design
Deep Learning Under Privileged Information Using Heteroscedastic Dropout
John Lambert, Ozan Sener, Silvio Savarese
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Use bounding-box privileged information to control heteroscedastic dropout variance | VGG-16 on ImageNet · ImageNet 75K | single-crop top-1 accuracy: 37.85 -> 42.30 | Improved |
| Use bounding-box privileged information to control heteroscedastic dropout variance | VGG-16 on ImageNet · ImageNet 200K | single-crop top-1 accuracy: 55.99 -> 55.51 | Hurt |
| Use bounding-box privileged information to control heteroscedastic dropout variance | VGG-16 on ImageNet · ImageNet 600K | single-crop top-1 accuracy: 66.66 -> 66.77 | No clear effect |
| Use heteroscedastic dropout with privileged images | LSTM machine translation · Multi-30K en→de | BLEU: 35.5 -> 38.4 | Improved |
| Use heteroscedastic dropout with privileged images | LSTM machine translation · Multi-30K de→en | BLEU: 40.19 -> 42.4 | Improved |
| Use heteroscedastic dropout rather than a fixed-variance Gaussian dropout | ImageNet classification · ImageNet 75K | single-crop top-1 accuracy: 38.80 -> 42.30 | Improved |
| Replace learned covariance with entries drawn from pure Gaussian noise (ablation) | Heteroscedastic dropout on image classification · ImageNet | accuracy: >3.8% | Hurt |
| Use only a small fraction of training examples with privileged information (ablation) | Heteroscedastic dropout for machine translation · Multi-30K | accuracy | Improved |
| Use only a small fraction of training examples with privileged information (ablation) | Heteroscedastic dropout for image classification · ImageNet | top-1 accuracy | Improved |
| Use adaptive validation-based learning-rate decay rather than fixed-interval decay (ablation) | Image classification models · ImageNet | accuracy: 5 - 10% | Improved |
| Remove information-bottleneck regularization (ablation) | Heteroscedastic dropout · Training | loss convergence | Hurt |
Closest papers by meaning
- Dropout with Expectation-linear Regularization (ICLR 2017 · similarity 0.87)
- Information Dropout: Learning Optimal Representations Through Noisy Computation (TPAMI 2018 · similarity 0.87)
- Variational Dropout Sparsifies Deep Neural Networks (ICML 2017 · similarity 0.86)
- Advanced Dropout: A Model-free Methodology for Bayesian Dropout Optimization (TPAMI 2021 · similarity 0.86)
- Contextual Dropout: An Efficient Sample-Dependent Dropout Module (ICLR 2021 · similarity 0.86)
- DRoP: Distributionally Robust Data Pruning (ICLR 2025 · similarity 0.84)
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.
