Research map / Learning theory and algorithms / Prediction and decision losses
Contrastive Neural Ratio Estimation
Benjamin Kurt Miller, Christoph Weniger, Patrick Forré
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 NRE-C instead of NRE-B (multiclass with bias term) | NRE · SBI benchmark, 10^3 simulations | C2ST: 0.867 -> 0.853 | Improved |
| Use NRE-C instead of NRE-B | NRE · SBI benchmark, 10^4 simulations | C2ST: 0.811 -> 0.762 | Improved |
| Use NRE-C instead of NRE-B | NRE · SBI benchmark, 10^5 simulations | C2ST: 0.762 -> 0.680 | Improved |
| NRE-C versus amortized NPE | NRE-C · SBI benchmark, 10^5 simulations | C2ST: 0.654 -> 0.680 | No clear effect |
| NRE-C versus NLE | NRE-C · SBI benchmark, 10^4 simulations | C2ST: 0.753 -> 0.762 | No clear effect |
| NRE-C versus sequential SNRE | NRE-C · SBI benchmark, 10^5 simulations | C2ST: 0.610 -> 0.680 | Hurt |
| Increase number of contrastive parameters K with unlimited data (ablation) | NRE-B and NRE-C · Draw (theta, x) every mini-batch, 3 SBI tasks | C2ST | Improved |
| Vary gamma with unlimited data (ablation) | NRE-C · Draw (theta, x) every mini-batch | C2ST | No clear effect |
| Use Large NN (128 units, 3 residual blocks) instead of Small NN (50 units, 2 blocks) (ablation) | NRE-C · SBI tasks, esp. SLCP | C2ST | Improved |
| Draw contrastive parameters from the prior each mini-batch (ablation) | NRE-B and NRE-C · 3 SBI tasks | C2ST sensitivity to K | Improved |
| Lower gamma with Large NN (ablation) | NRE-C · Draw theta / fixed data regimes | C2ST | No clear effect |
| Increase K with fixed gamma (ablation) | NRE-C · SLCP, mutual information bound | -I_w^(0) bound | Mixed |
Closest papers by meaning
- Learning Gaussian Conditional Distributions using Neural Ratio Estimation is Hard (NeurIPS 2026 · similarity 0.87)
- Direct Amortized Likelihood Ratio Estimation (AAAI 2024 · similarity 0.86)
- Sequential Neural Score Estimation: Likelihood-Free Inference with Conditional Score Based Diffusion Models (ICML 2024 · similarity 0.85)
- CoLT: The conditional localization test for assessing the accuracy of neural posterior estimates (NeurIPS 2025 · similarity 0.84)
- What Does It Take to Build a Performant Selective Classifier? (NeurIPS 2025 · similarity 0.84)
- Crossing the Validation crisis: Cross-validation reduces benchmarking variance surprisingly well (NeurIPS 2026 · similarity 0.83)
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.
