Interpretable Human-Label-Free Deep Learning for Real-Bogus Classification with Uncertainty Quantification
Astronomy ML: a dual-network classifier separates real from bogus telescope transients using injected simulations instead of human labels. Off-topic for agent builders; the label-free training recipe is the takeaway.
**The gist** An astronomy paper: a **dual-network** classifier separates real telescope transients from bogus detections, trained with **asymmetric co-teaching** on **simulated injections** plus contaminated survey data — no human labels. A hybrid of **MC dropout and deep ensembles** delivers calibrated uncertainty at lower cost than full ensembles.
**Why it matters** Outside the agent-engineering beat, but the recipe travels: when labels are costly, injecting synthetic positives into noisy real negatives can still train a robust triage classifier, and pairing two networks doubles as cheap, calibrated **uncertainty quantification** for any automated filtering pipeline.
**The gist** An astronomy paper: a **dual-network** classifier separates real telescope transients from bogus detections, trained with **asymmetric co-teaching** on **simulated injections** plus contaminated survey data — no human labels. A hybrid of **MC dropout and deep ensembles** delivers calibrated uncertainty at lower cost than full ensembles. **Why it matters** Outside the agent-engineering beat, but the recipe travels: when labels are costly, injecting synthetic positives into noisy real negatives can still train a robust triage classifier, and pairing two networks doubles as cheap, calibrated **uncertainty quantification** for any automated filtering pipeline. **Watch out** The abstract gives **no headline accuracy numbers**; single-source identification stays limited by ambiguous light-curve labels, and moving to a new survey means re-running the entire **injection-based training pipeline**.