Explain the stronger response to S-shaped signals than to dispersed bursts

Explain why the EfficientNet-B0 CNN technosignature detector produces a stronger Grad-CAM response to injected S-shaped signals than to dispersed-burst signals, despite the dispersed-burst signals depositing substantially more energy.

Background

The detector was evaluated on several synthetic technosignature morphologies injected into real Green Bank backgrounds. The S-shaped family achieved high recall and produced a larger total Grad-CAM perturbation, whereas the dispersed-burst family was almost always missed and produced a weaker response despite depositing approximately thirteen times as much energy as the pulsed family. The authors tested curvature, response concentration, and injected energy as possible explanations, but these factors did not account for the difference.

The unresolved issue is what learned feature or interaction within the frozen CNN causes the markedly different responses to these two nonstandard signal morphologies. Resolving it could clarify whether the detector’s failures arise primarily from morphology-specific training coverage or from the way the network weights features that it does register.

References

D1-5 deposits about 13$\times$ the energy of D1-4 yet provokes a weaker response, 0.29$\times$ the total CAM perturbation of D1-1 against 0.47$\times$ for D1-4, and this analysis does not explain why.

— Missed Detections and False Alarms in a CNN Technosignature Detector on Breakthrough Listen Data  (2610.22050 - Imran, 15 Oct 2026) in Section 4.3, subsection “What the ON-only input explains”

Two explanations remain: D2-1 is not tilted enough to match the signals it learned, or vertical lines are common interference in the training negatives and it learned to reject them. These were not separated here.

— Missed Detections and False Alarms in a CNN Technosignature Detector on Breakthrough Listen Data  (2610.22050 - Imran, 15 Oct 2026) in Section 5.2, subsection “Rejection by appearance, not origin”