Quantifying uncertainty in deep learning approaches to radio galaxy classification
Problem
Deep learning classifiers of radio galaxies provide predictions without reliable uncertainty estimates, making their outputs hard to trust.
Why it matters
Trustworthy automated classification is needed to handle large astronomical radio survey datasets.
What the AI did
Variational inference was used to quantify posterior uncertainty in deep learning radio galaxy classification, testing weight priors and pruning strategies.
Results
Model posterior variance correlated with human labelling uncertainty; 30 per cent of fully connected layer weights pruned without significant performance loss; sparse prior gave better-calibrated uncertainty.
Impact
Evidence supports improved uncertainty-aware classification, though calibration effects and cold posterior effect remain.
Record
| Venue | Monthly Notices of the Royal Astronomical Society |
|---|---|
| Published | 2022 |
| Work type | article |
| Authors | Devina Mohan, Anna M. M. Scaife, Fiona Porter, Mike Walmsley, Micah Bowles |
| Citations | 16 (OpenAlex, at import time) |
| References | 52 |
| Model or method | Bayesian deep learning with variational inference |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1093/mnras/stac223 |
| Open access | hybrid |
| Keywords | deep learning, uncertainty quantification, radio galaxies, Bayesian neural networks, pruning |
Importance rating
incremental / niche
Uncertainty quantification for radio galaxy classifiers is a methodological advance for trustworthy survey processing; no scale stated, niche within astronomy.
More in Physics & Astronomy
| # | Paper | Year | Citations | Importance |
|---|---|---|---|---|
| 1 | Magnetic control of tokamak plasmas through deep reinforcement learning | 2022 | 739 |
7
|
| 2 | Solving the quantum many-body problem with artificial neural networks | 2017 | 2,205 |
7
|
| 3 | Learning high-accuracy error decoding for quantum processors | 2024 | 90 |
7
|
| 4 | Optimizing Quantum Error Correction Codes with Reinforcement Learning | 2019 | 158 |
7
|
| 5 | Machine-learning-assisted correction of correlated qubit errors in a topological code | 2018 | 141 |
7
|