Machine Learning Reveals Unknown Transient Phenomena in Historic Images

· ai history math · Source ↗

TLDR

  • ML classifier (AUC 0.81) applied to 107,875 pre-Sputnik observatory plate transients confirms two anomalous correlations survive artifact filtering: a nuclear-test window and an Earth-shadow deficit.

Key Takeaways

  • Model trained on 250 expert-labeled image pairs (real vs. plate defect); out-of-fold AUC 0.81, sensitivity 0.71, specificity 0.71 – meaningful but not high-confidence discrimination.
  • After ML artifact removal, transient counts remain elevated within ±1 day of nuclear tests (p=.024 overall; p<.0001 for highest-probability real transients).
  • Shadow deficit – fewer transients when the observation point is in Earth’s shadow – holds at p<.0001 and is strongest in the highest-probability transient tier.
  • The core methodological contribution is replacing expert-only visual review with a deployable ML pipeline that scales to ~108K archived observations.
  • Prior skeptical claim (automated pipelines detect only plate defects) is now partially rebutted; high-probability real events still show the nuclear window and shadow patterns.

Hacker News Comment Review

  • Leading technical objection: nuclear tests produce high-energy radiation that directly fogs photographic film, which would trivially explain the nuclear-window correlation without invoking novel sky objects.
  • The paper itself acknowledges it cannot falsify hypotheses including pre-Sputnik satellite launches or non-human technosignatures – commenters flag this caveat as the authors hedging toward speculation the data cannot support.
  • Thread is short; no engagement with ML methodology specifics (architecture, training data provenance, out-of-fold split) – skepticism focuses on physical confounds, not model validity.

Notable Comments

  • @alternator: argues film radiation exposure from nearby nuclear tests is the parsimonious explanation; protective housing would not always help given plate handling timelines.
  • @surprisetalk: surfaces the paper’s own language – “non-human technosignature” and pre-Sputnik launches are listed as unfalsifiable hypotheses given data limitations.

Original | Discuss on HN