Research map / Label-efficient and robust learning / Recognition and long-tail
Scaling up Data Augmentation MCMC via Calibration
Leo L. Duan, James E. Johndrow, David B. Dunson
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 |
|---|---|---|---|
| add calibration (CDA) to data augmentation | DA Gibbs sampler · Co-browsing hierarchical binomial (n=59,792) | effective sample size per iteration: 0.0085 -> 0.5013 | Improved |
| add calibration (CDA) to data augmentation | DA Gibbs sampler · Co-browsing hierarchical binomial (n=59,792) | computing time per effective sample: 140.4 sec -> 0.48 sec | Improved |
| use CDA instead of HMC | hierarchical binomial model · Co-browsing hierarchical binomial (n=59,792) | computing time per effective sample: 1.3 sec -> 0.48 sec | Improved |
| use CDA instead of HMC | hierarchical binomial model · Co-browsing hierarchical binomial (n=59,792) | effective sample size per iteration: 0.5013 vs 0.8404 | Hurt |
| add calibration (CDA) to data augmentation | DA Gibbs sampler · Co-browsing Poisson log-normal (n=59,792) | prediction RMSE: 33.21 -> 8.52 | Improved |
| use CDA instead of HMC | Poisson log-normal model · Co-browsing Poisson log-normal (n=59,792) | prediction RMSE: 13.18 -> 8.52 | Improved |
| add calibration (CDA) to data augmentation | DA Gibbs sampler · Co-browsing Poisson log-normal (n=59,792) | computing time per effective sample: 346.4 sec -> 11.5 sec | Improved |
| use CDA instead of HMC | Poisson log-normal model · Co-browsing Poisson log-normal (n=59,792) | computing time per effective sample: 3240.6 sec -> 11.5 sec | Improved |
| increase r from 1 to the optimal 1,000 in intercept-only probit (ablation) | CDA without M-H · Intercept-only probit (n=1e4, one success) | autocorrelation | Improved |
| increase r from 1,000 to 5,000 in intercept-only probit (ablation) | CDA without M-H · Intercept-only probit (n=1e4, one success) | autocorrelation | No clear effect |
| add CDA to subsampled Polya-Gamma | subsampling DA · Logistic regression, n=1e5 to 1e8 | effective sample size | Improved |
| use CDA instead of M-H with scaled DA proposal or normal proposal | logistic rare event · Logistic regression rare event (n=1e5, 50 events) | M-H acceptance rate: acceptance rate of 0.8 | Improved |
Closest papers by meaning
- Divide et Calibra: Multiclass Local Calibration via Vector Quantization (NeurIPS 2026 · similarity 0.78)
- Semantic-Aware Data Augmentation for Text-to-Image Synthesis (AAAI 2024 · similarity 0.78)
- UtilGen: Utility-Centric Generative Data Augmentation with Dual-Level Task Adaptation (NeurIPS 2025 · similarity 0.78)
- NSNQuant: A Double Normalization Approach for Calibration-Free Low-Bit Vector Quantization of KV Cache (NeurIPS 2025 · similarity 0.78)
- A Benchmark Study on Calibration (ICLR 2024 · similarity 0.77)
- Thermometer: Towards Universal Calibration for Large Language Models (ICML 2024 · similarity 0.77)
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.
