Research map / Learning theory and algorithms / Kernels and regression theory
Data Thinning for Convolution-Closed Distributions
Anna Neufeld, Ameer Dharamshi, Lucy L. Gao, Daniela Witten
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 data thinning with eps_train=0.8 instead of 0.5 for binomial PCA rank selection (ablation) | data thinning loss curve · Binomial PCA simulation (K*=10) | correct selection of K*: eps_train=0.5 fails to select K*=10; eps_train=0.8 selects it | Improved |
| increase eps_train toward 1 (ablation) | single-fold data thinning · all three simulations | proportion selecting true K* | Hurt |
| use data thinning instead of naive reuse of data for loss computation | K selection by loss minimization · binomial PCA and gamma clustering simulations | loss curve shape: sharp minima vs monotonically decreasing curves | Improved |
| use multifold thinning (M=5) instead of single-fold thinning (eps_train=0.8) (ablation) | data thinning K selection · binomial PCA and small gamma simulations | frequency of selecting true K | Improved |
| use multifold thinning (M=5) instead of single-fold thinning (ablation) | data thinning K selection · gamma clustering (d=100, K*=10) | frequency of selecting true K | No clear effect |
| replace negative log-likelihood loss with mean squared error loss (ablation) | data thinning K selection · binomial PCA and gamma clustering simulations | selection of K* | No clear effect |
| data thinning instead of sample splitting | forward stepwise selection then inference · n=100, p=20 iid regression, eps=0.2 and 0.8 | detection and power | No clear effect |
| data thinning instead of sample splitting | forward stepwise selection then inference · n=42 with one high-leverage point | detection and power | Improved |
| thin with too small a nuisance variance (sigma tilde squared < sigma squared) (ablation) | Gaussian data thinning · X~N(7,5), eps=0.44 | correlation between X1 and X2 | Hurt |
| use data thinning with eps=0.5 for PC selection instead of in-sample SSE | Seurat PCA · PBMC scRNA-seq | loss versus K | Improved |
Closest papers by meaning
- Low-Rank Thinning (ICML 2025 · similarity 0.87)
- Supervised Kernel Thinning (NeurIPS 2024 · similarity 0.84)
- DRoP: Distributionally Robust Data Pruning (ICLR 2025 · similarity 0.84)
- $\mathbb{D}^2$ Pruning: Message Passing for Balancing Diversity & Difficulty in Data Pruning (ICLR 2024 · similarity 0.83)
- Crossing the Validation crisis: Cross-validation reduces benchmarking variance surprisingly well (NeurIPS 2026 · similarity 0.83)
- Debiased Distribution Compression (ICML 2024 · 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.
