Tutorial 05 of 09
Classification
Gate: braai CNN
Type: ALeRCE

REAL, BOGUS, THEN WHAT.

Nothing downstream is trusted until a real/bogus gate scores it, and nothing earns a type label until it clears that gate. This tutorial covers the pretrained CNN that does the grading, and the two ways the pipeline guesses what a real detection actually is. Nine tutorials, read in order — this is the fifth.

Gate
braai (pretrained CNN)
Native recall
0.957
Own-cutout recall
0.05–0.14
Types
AGN, SN, VS, asteroid

Unlike the Overview page, nothing on this page is invented. This describes the real pipeline — the same one the viewer, survey and cell browsers actually run.

— 05 . REAL, BOGUS, AND TYPE

A pretrained CNN grades every detection, then a second pass guesses what it is.

Nothing downstream is trusted until a real/bogus gate scores it, and nothing earns a type label until it clears that gate.

The gate isn't judging the object — it's judging the image.

In this project

braai, the real/bogus gate

braai is a pretrained CNN published by Caltech/ZTF, used as-is with no local training. It scores each detection's science/reference/difference triplet with one number, P(real), from 0 to 1. On ZTF's own clean stamps for known real objects it recovers 0.957 of them — a trustworthy, well-validated classifier. On this project's own homemade cutouts it reached only 0.05–0.14 recall.

Picky, not broken. That gap traced to one documented cause: a sub-pixel registration "dipole" baked into the homemade cutouts — not a flaw in braai, not a bug in the gate. A classifier refusing a smeared image is doing its job. Strictness about image quality is correct behavior, not failure.

Passing the gate answers what happened; typing answers what it was.

Stamp one image, one guess

An older TensorFlow-1-era ALeRCE CNN, run locally on a single cutout, sorts survivors into AGN, supernova, variable star, or asteroid.

Light curve history over time

ALeRCE's Balanced Hierarchical Random Forest reads brightness-vs-time across 15 fine classes grouped under three top-level categories — Periodic, Stochastic, Transient.

In this project

114 detections, one field, one night

A single 2020 field run split 45 bogus, 36 variable star, 12 AGN, 21 uncertain — a sensible spread for an ordinary equatorial star field. The stamp classifier scored ~0.70 overall accuracy on a fair labelled evaluation, weakest at telling AGN from variable stars apart (~0.55) — both are unresolved point sources at a fixed position, a genuine astrophysical ambiguity, not a bug. The light-curve classifier was strongest on periodic stars — eclipsing binaries 5 of 5 correct, RR Lyrae 3 of 5, on a 29-object set — for 0.66 top-level and 0.38 fine-class accuracy overall.

The whole point of a tutorial

Go watch it run.

Everything above is the real pipeline, explained in the order it runs. The browsers below are that same pipeline's actual output — a working ZTF difference-imaging system, not a mockup of one.

Or read what this project actually is