Electromagnetic signal tasks stop needing one expert each: GLCR routes at 96.16% accuracy with a two-stage design
Synopsis
Electromagnetic signal processing has long trained a separate expert per task, limiting cross-task reuse, while electromagnetic foundation models pay for large-scale pretraining and shared-backbone inference. GLCR instead routes over a frozen heterogeneous expert pool: a lightweight one-dimensional CNN with an additive-margin head makes the global dispatch, and a local re-evaluation with four physical statistics handles confusable task pairs. On the HeteroSignal-7 benchmark it reaches 96.16% routing accuracy and a 76.30% system score, above all 19 baselines, at 1.43 ms routing latency per sample.
Interpretation
Multi-task electromagnetic signal processing can now run one lightweight router plus one selected frozen expert, instead of training a model per task or sharing one backbone across all tasks. The two prior routes each carried a cost: single-task experts were developed independently with fixed inputs and objectives and transferred poorly across tasks, while electromagnetic foundation models covered more tasks but incurred substantial pretraining and shared-backbone inference costs. On the cleaned HeteroSignal-7 test set, GLCR reaches 96.16% routing accuracy and a 76.30% system score, 1.16% and 1.14% above the respective second-best baselines, and records the lowest routing latency among compared methods at 1.43 ms per sample.
Routing runs in two steps: a fast global decision first, and only samples landing on a confusable task pair go to a second-stage re-evaluation. Single-stage global routing mixes up similar time-domain waveforms from different tasks at low SNR, concentrating errors on particular task pairs. The first stage is a four-layer one-dimensional CNN with an additive-margin head producing seven routing scores; the second is a histogram-based gradient-boosted tree over four physical statistics, namely noncircularity magnitude, spectral centroid, spectral flatness and amplitude kurtosis, triggered only when the top two candidates match a confusion pair accounting for more than 10% of validation errors, which in this benchmark is only the RML2018.01a and SD-BENCH3 pair.
Perspective
The demonstrated scope is the HeteroSignal-7 benchmark: seven datasets, six task types and 140,000 samples paired with seven frozen classification and waveform-recovery experts, with routing limited to closed-set task selection; only FS-SEI among the seven datasets comes from real-world measurements.
The system score remains below the perfect-routing upper bound of 78.46%, and many remaining errors concentrate on the RML2018.01a and SD-BENCH3 confusion pair, indicating the four statistics do not capture all physical differences between the two classes; the local re-evaluator handles only one confusion pair, and open-set routing and more measured signal sources are left as future work.