Research map / Label-efficient and robust learning / Semi-supervised learning
VisRec: A Semi-Supervised Approach to Visibility Data Reconstruction in Radio Astronomy
Read the paperCheck your idea against 108,967 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.
Proposes VisRec, a model-agnostic semi-supervised framework that combines augmented supervised training on few labeled visibility samples with consistency training on unlabeled visibilities to reconstruct radio interferometric visibility data.
Problem
Deep learning visibility reconstruction in radio astronomy needs large labeled datasets (clean ground-truth images) that are costly to obtain, and supervised methods are fragile under noise, missing data, and telescope configuration shifts.
Approach
A model-agnostic semi-supervised scheme pairing label-invariant and label-variant uv-plane augmentations on labeled data with a consistency loss that makes reconstructions of corrupted unlabeled visibilities match the reconstruction of the clean unlabeled input (used as pseudo-label).
Key results
- On the EHT dataset, CNN VisRec achieves PSNR 23.37 dB and SSIM 0.892 versus 22.22 dB / 0.864 for supervised CNN (Radionet).
- NF VisRec with only 2k labeled examples performs on par with supervised NF trained on over 16k labeled examples in ideal conditions.
- At 50% missing visibility points, NF VisRec's PSNR drops by only about 1% versus about 10% for supervised NF.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add VisRec semi-supervised consistency training to Neural Field | Neural Field · EHT Dataset | PSNR (dB): 21.36 -> 22.63 | Helped |
| add VisRec semi-supervised consistency training to Neural Field | Neural Field · VLBA Dataset | PSNR (dB): 20.26 -> 22.87 | Helped |
| add VisRec semi-supervised consistency training to CNN | CNN · EHT Dataset | PSNR (dB): 22.22 -> 23.37 | Helped |
| add VisRec semi-supervised consistency training to CNN | CNN · VLBA Dataset | PSNR (dB): 22.15 -> 23.94 | Helped |
| add supervised visibility augmentations (ablation) | Neural Field supervised training · EHT Dataset | PSNR (dB): 21.36 -> 21.89 | Helped |
| add supervised visibility augmentations (ablation) | Neural Field supervised training · VLBA Dataset | PSNR (dB): 20.26 -> 21.24 | Helped |
| add supervised visibility augmentations (ablation) | CNN supervised training · EHT Dataset | PSNR (dB): 22.22 -> 22.32 | No clear effect |
| add supervised visibility augmentations (ablation) | CNN supervised training · VLBA Dataset | PSNR (dB): 22.15 -> 23.21 | Helped |
| replace supervised pathway with self-supervised training (ablation) | Neural Field VisRec · EHT Dataset | PSNR (dB): 22.63 -> 11.45 | Hurt |
| replace supervised pathway with self-supervised training (ablation) | Neural Field VisRec · VLBA Dataset | PSNR (dB): 22.87 -> 13.29 | Hurt |
| replace supervised pathway with self-supervised training (ablation) | CNN VisRec · EHT Dataset | PSNR (dB): 23.37 -> 15.18 | Hurt |
| replace supervised pathway with self-supervised training (ablation) | CNN VisRec · VLBA Dataset | PSNR (dB): 23.94 -> 14.65 | Hurt |
Surprising findings
- Self-supervised variant (Noise2Noise-style) performs worse than dirty image in several metrics, and the VisRec variant without supervised augmentation sometimes matches or beats full VisRec on certain metrics.
Limitations
- Evaluated only on simulated visibility data from two telescope configurations (EHT and VLBA).
- Generalization across datasets still shows notable performance drops for all models.
- Augmentation and training approach may not transfer directly to other domains such as MRI or seismic imaging.
Experimental setup
Datasets: Galaxy10 DECals EHT-simulated visibility dataset (eht-imaging) VLBA-simulated visibility dataset
Benchmarks: EHT visibility dataset VLBA visibility dataset
Metrics: LFD SSIM PSNR
Baselines: CLEAN Noise2Astro Neural Field (Neural Interferometry) CNN (Radionet) NF Supervised + Aug NF Self-Supervised CNN Self-Supervised
Builds on: Neural Interferometry Radionet Noise2Noise CLEAN
Closest papers by meaning
- PolarRec: Improving Radio Interferometric Data Reconstruction Using Polar Coordinates (CVPR 2024 · similarity 0.88)
- A systematic evaluation of vision-language models for observational astronomical reasoning tasks (NeurIPS 2026 · similarity 0.82)
- Learning What's Real: Disentangling Signals and Measurement Artifacts in Multi-Sensor Data, with Applications to Astrophysics (NeurIPS 2026 · similarity 0.81)
- AstroVisBench: A Code Benchmark for Scientific Computing and Visualization in Astronomy (NeurIPS 2025 · similarity 0.81)
- Galileo: Learning Global & Local Features of Many Remote Sensing Modalities (ICML 2025 · similarity 0.81)
- BERTground: A Transformer-Based Model of Background Spectra on the ISS-Based NICER Space Telescope (AAAI 2024 · 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.