Research map / Learning theory and algorithms / Matrix and tensor methods
Can Metadata Fix the Gauge? Calibration Turns Sparse Multi-Source Learning from Sparse PCA into Sparse Mean Recovery
Yibo Zhou, Yirong Xiang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that source-specific sign gauges make pooled multi-source sparse learning reduce to sparse PCA, and that task-independent relative-gauge metadata lets synchronization restore a sparse mean-recovery problem.
Problem
Sparse Gaussian multi-source learning where each source orients its coordinates with an unknown sign gauge; without calibration, pooling erases the mean signal and the problem becomes sparse PCA.
Approach
Uses task-independent relative-gauge metadata with synchronization to align source-specific sign gauges, turning the problem from sparse PCA into sparse mean recovery.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add valid relative-gauge calibration (synchronization) before pooling | Sparse multi-source recovery · Digits (imposed gauges) | accuracy: 0.500 -> 0.968 | Improved |
| Add valid relative-gauge calibration (synchronization) before pooling | Sparse multi-source recovery · Breast cancer (imposed gauges) | accuracy: 0.501 -> 0.932 | Improved |
| Use random metadata instead of valid calibration (ablation) | Sparse multi-source recovery · Digits | accuracy: 0.500 -> 0.525 | No clear effect |
| Use shuffled metadata instead of valid calibration (ablation) | Sparse multi-source recovery · Breast cancer | accuracy: 0.501 -> 0.557 | No clear effect |
| Use noisy absolute calibration with 10% flips instead of relative synchronization (ablation) | Sparse multi-source recovery · Digits | accuracy: 0.968 -> 0.874 | Hurt |
| Use source-label task alignment (uses downstream labels) instead of calibration (ablation) | Sparse multi-source recovery · Digits | accuracy: 0.968 -> 0.818 | Hurt |
| Reuse one synchronized calibration graph across 32 tasks | GSL · Reuse experiment | Corr: 0.928 -> 0.928 | No clear effect |
| Increase aligned source count at fixed aligned mass from 1 to 32 (ablation) | Calibrated recovery · Synthetic, M_q=4096 | Corr: 0.632 -> 0.917 | Improved |
| Increase aligned source count from 32 to 128 (ablation) | Calibrated recovery · Synthetic, M_q=4096 | Corr: 0.917 -> 0.916 | No clear effect |
| Use source-label PCA instead of calibration (ablation) | Sparse multi-source recovery · Reuse experiment | Corr: 0.928 -> 0.040 | Hurt |
Surprising findings
- Pooling across sources without calibration erases the mean signal, reducing the source-labelled experiment to sparse PCA.
Experimental setup
Builds on: sparse PCA
Closest papers by meaning
- Propensity Score Alignment of Unpaired Multimodal Data (NeurIPS 2024 · similarity 0.83)
- Beyond One-Hot Labels: Semantic Mixing for Model Calibration (ICML 2025 · similarity 0.81)
- Calibration without labels in multiple testing (NeurIPS 2026 · similarity 0.81)
- The Price of Sparsity: Sufficient Conditions for Sparse Recovery using Sparse and Sparsified Measurements (NeurIPS 2025 · similarity 0.81)
- PCA++: How Uniformity Induces Robustness to Background Noise in Contrastive Learning (NeurIPS 2025 · similarity 0.81)
- Sparse components distinguish visual pathways & their alignment to neural networks (ICLR 2025 · similarity 0.81)
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.
