A technique that decomposes predictions into independent, complementary factors learned through separate pathways then recombined.