Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Dropout Inference in Bayesian Neural Networks with Alpha-divergences
Read the paperCheck 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 |
|---|---|---|---|
| Use α=0.5 instead of α=0 for dropout inference (ablation) | Dropout BNN regression · UCI regression datasets | negative test log likelihood: Boston: 2.42 ± 0.05 -> 2.38 ± 0.06; Concrete: 2.98 ± 0.02 -> 2.88 ± 0.02; Energy: 1.75 ± 0.01 -> 0.74 ± 0.02; Kin8nm: -0 | Helped |
| Use α=0.5 instead of α=0 for dropout inference (ablation) | Dropout BNN regression · Power and Yacht | negative test log likelihood: Power: 2.79 ± 0.01 -> 2.78 ± 0.01; Yacht: 1.38 ± 0.01 -> 1.08 ± 0.04 | Helped |
| Use α=1.0 instead of α=0 for dropout inference (ablation) | Dropout BNN regression · Protein, Red wine, Naval, and Year | negative test log likelihood: Protein: 2.87 ± 0.00 -> 2.86 ± 0.00; Red wine: 0.92 ± 0.01 -> 0.95 ± 0.02; Naval: -2.80 ± 0.00 -> -2.80 ± 0.00; Year: 3. | Mixed |
| Use α=0.5 instead of α=0 for dropout inference (ablation) | Dropout BNN regression · UCI regression datasets | test RMSE: Concrete: 4.92 ± 0.13 -> 4.62 ± 0.12; Boston: 2.85 ± 0.19 -> 2.97 ± 0.19; Energy: 1.02 ± 0.03 -> 1.11 ± 0.02 | Mixed |
| Use α=0.5 instead of α=0 for dropout inference (ablation) | Dropout BNN · MNIST fully connected classification | test RMSE | Helped |
| Use α=0.5 instead of α=0 for dropout inference (ablation) | Dropout BNN · MNIST fully connected classification | test log likelihood | Helped |
| Increase training Monte Carlo samples from K=1 to K=10 (ablation) | Dropout BB-α (α=1) · MNIST, fully connected NN | test log likelihood: -600 within 1000 seconds; -500 after 3000 seconds | Helped |
| Increase training Monte Carlo samples from K=1 to K=10 (ablation) | Dropout BB-α (α=1) · MNIST, fully connected NN | test accuracy: 98.8% | No clear effect |
| Increase training Monte Carlo samples from K=10 to K=100 (ablation) | Dropout BB-α (α=1) · MNIST, fully connected NN | test log likelihood: -500 after 3000 seconds | Hurt |
| Use dropout-trained Bayesian inference instead of a deterministic maximum-likelihood network | MNIST fully connected network · MNIST adversarial examples | predictive uncertainty | Helped |
Closest papers by meaning
- Variational Bayesian dropout: pitfalls and fixes (ICML 2018 · similarity 0.91)
- Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning (ICML 2016 · similarity 0.91)
- Variational Dropout Sparsifies Deep Neural Networks (ICML 2017 · similarity 0.90)
- Structured Dropout Variational Inference for Bayesian Neural Networks (NeurIPS 2021 · similarity 0.89)
- Contextual Dropout: An Efficient Sample-Dependent Dropout Module (ICLR 2021 · similarity 0.88)
- Maxitive Donsker-Varadhan Formulation for Possibilistic Variational Inference (NeurIPS 2026 · similarity 0.86)
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.