Test whether the observed failures persist across retraining

Determine whether the missed detections, false alarms, and other failure patterns of the frozen EfficientNet-B0 CNN persist across retraining with different random seeds and validation folds.

Background

All experiments used one training run, one random seed, and one validation fold. Consequently, the reported behavior may reflect properties of the particular fitted network rather than stable properties of the competition baseline architecture or training procedure.

A systematic retraining study would establish whether the morphology-dependent recall failures, non-monotonic amplitude response, and high false-positive rate for high-drift interference are reproducible. The paper explicitly leaves this persistence question untested.

References

The detector was trained once, with one random seed and one validation fold, so the failures reported here belong to that training run; whether they persist across retraining was not tested.

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