Determine the cause of recall decline at high signal amplitude

Determine why the EfficientNet-B0 technosignature detector’s recall declines as injected signal amplitude increases, including whether post-injection normalization or a learned association between brightness and interference causes the decline.

Background

The detector’s recall decreases at high amplitudes for several genuine-signal families, including the narrowband-like calibration needle on which it was trained. The effect cannot be explained by injected values exceeding the normal input range. The authors identify two unresolved explanations: the experimental injections were not renormalized in the same way as the competition’s training data, or the training data may have associated bright signals with interference.

Renormalizing each injected sample before scoring would test the first explanation, while additional analyses of brightness and class-conditional training distributions would be needed to assess the second. The paper therefore treats the mechanism behind the amplitude effect as unresolved.

References

The fall in recall at high amplitude is also unexplained. A value-range artefact can be ruled out: the data reach values of about 37, with a 99.9th percentile near 12.6, so an injected line of amplitude 4 is well inside the normal range. Two hypotheses remain untested.

— Missed Detections and False Alarms in a CNN Technosignature Detector on Breakthrough Listen Data  (2610.22050 - Imran, 15 Oct 2026) in Section 6, “Limitations”