| _version_ | 1866901607043563520 |
|---|---|
| author | Cakir, Dorukhan Hasanova, Asmar |
| author_facet | Cakir, Dorukhan Hasanova, Asmar |
| contents | <p>Deep learning has pushed Automatic Modulation Classification (AMC, receiver-side blind modulation recognition) to high accuracy, yet the reliability of predicted confidence scores has received little attention. A classifier that reports 85% confidence at low signal-to-noise ratio (SNR) while achieving only 40% accuracy creates serious risks for cognitive radio, spectrum sharing, and electronic warfare decision systems. To our knowledge, we present the first systematic study of confidence calibration in deep AMC classifiers. On RadioML 2016.10a with 5- seed replication, we find that the optimal calibration temperature decreases monotonically with SNR (Pearson r = −0.918±0.013), transitioning from overconfidence at low SNR to underconfidence at high SNR. A single global temperature cannot correct this opposing behavior. Our method fits a separate temperature parameter for each SNR operating point on a held-out validation set and applies it at inference with zero retraining.</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_19747126 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | Calibrating Deep AMC Classifiers with SNR-Adaptive Temperature Scaling Cakir, Dorukhan Hasanova, Asmar <p>Deep learning has pushed Automatic Modulation Classification (AMC, receiver-side blind modulation recognition) to high accuracy, yet the reliability of predicted confidence scores has received little attention. A classifier that reports 85% confidence at low signal-to-noise ratio (SNR) while achieving only 40% accuracy creates serious risks for cognitive radio, spectrum sharing, and electronic warfare decision systems. To our knowledge, we present the first systematic study of confidence calibration in deep AMC classifiers. On RadioML 2016.10a with 5- seed replication, we find that the optimal calibration temperature decreases monotonically with SNR (Pearson r = −0.918±0.013), transitioning from overconfidence at low SNR to underconfidence at high SNR. A single global temperature cannot correct this opposing behavior. Our method fits a separate temperature parameter for each SNR operating point on a held-out validation set and applies it at inference with zero retraining.</p> |
| title | Calibrating Deep AMC Classifiers with SNR-Adaptive Temperature Scaling |
| url | https://doi.org/10.5281/zenodo.19747126 |