Research map / Label-efficient and robust learning / Recognition and long-tail

Robust Automatic Modulation Classification with Fuzzy Regularization

ICML 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Learning from unlabeled and positive-unlabeled data

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 Fuzzy Regularization (FR), an adaptive-gradient penalty on prediction ambiguity added to deep AMC models, improving accuracy and noise robustness on RadioML datasets.

Problem

Automatic modulation classification suffers from prediction ambiguity between similar modulation classes (near-identical top-class probabilities), which degrades reliability, especially at low SNR and under noise.

Approach

A regularization term computed from the normalized L2-norm-based ambiguity of the top-k predicted probabilities, with a log-normal batch-level adaptive scaling so the gradient grows with ambiguity and shrinks as predictions sharpen; added as a plug-in to existing AMC models.

Key results

• FR improves ThreeStream on Data2018 from 82.08% to 86.75% ACC (H-ACC 93.74% to 96.62%) and MCLDNN on Data2016a from 69.66% to 73.79% ACC.
• On Noise2016b at 60% noise factor, MCLDNN+FR reaches 74.31% ACC vs 59.02% for MCLDNN alone.
• Gains across five models and three datasets are reported as typically 1-4% ACC; FR gains are larger when the fraction of ambiguous samples is larger.

Measured effects reported in the paper

Surprising findings

• Improvement from FR varies across datasets; the strongest baseline on Data2016b saw only about 0.3-0.4% ACC gain, attributed to an already near-optimal classification plane.

Limitations

• The theoretical modeling assumes binary classification with fixed per-batch accuracy of 0.5, a strong simplification.
• Hyperparameters gamma and k require per-task tuning; the optimal k depends on the number of semantically similar classes.
• Gains are modest on some datasets (e.g., Data2016b, FEAT with some noise settings show negligible or negative change).

Experimental setup

Datasets: RadioML 2016.10a RadioML 2016.10b RadioML 2018.01A

Benchmarks: RadioML 2016.10a RadioML 2016.10b RadioML 2018.01A

Metrics: F1-Score ACC H-ACC

Baselines: DAELSTM FEAT MCLDNN ThreeStream ResNet

Builds on: label smoothing curriculum learning

Closest papers by meaning

• Content-aware Balanced Spectrum Encoding in Masked Modeling for Time Series Classification (AAAI 2025 · similarity 0.78)
• IncepSeqNet: Advancing Signal Classification with Multi-Shape Augmentation (Student Abstract) (AAAI 2024 · similarity 0.78)
• Label Noise: Ignorance Is Bliss (NeurIPS 2024 · similarity 0.77)
• Adaptive Estimation and Learning under Temporal Distribution Shift (ICML 2025 · similarity 0.76)
• CATCH: Channel-Aware Multivariate Time Series Anomaly Detection via Frequency Patching (ICLR 2025 · similarity 0.76)
• An Optimal Transport-based Latent Mixer for Robust Multi-modal Learning (AAAI 2025 · similarity 0.76)