Hierarchical mixture of experts

Identity Notes & Concept Scope The inspected Jordan–Jacobs construction uses generalized-linear experts and softmax gates with shared external input, with a finite tree defining nested soft partitions. It is static conditional modeling, distinct from temporal hidden Markov decision trees, hard classification trees, hierarchical clustering and language-specific sparse MoE routing. A mixture density network shares conditional-mixture semantics but need not impose this tree factorization or generalized-linear leaf structure. Flattening a fixed tree gives leaf weights without making every flat gating parameterization the same hierarchical model. EM, IRLS and the online surrogate are fitting procedures, not the model identity. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 382 preceding native IDs/names/aliases for the four researched candidates. Both full-checkout fixed-string searches included ignored and hidden files, with 45 and 45 matching lines. All 31 unique matched contents were read: prior specialized inference records, source/review context and pending leads, not exact records for the three new identities. Expanded titles, source IDs, plural forms and complete/partial-factorizability terminology were included. All 68 semantic label/synonym hits and 23 exact fully annotated AIO CLASS nodes with outgoing edges were read. Their generated YAMLs and seven native neighbors were read or byte-verified unchanged after prior complete reading, totaling 30 contextual records. The broader phrase structured variational inference is searched but not asserted as an unconditional alias of structured mean field. Boltzmann trees remains an unresolved full-text lead. Absence is bounded to the pinned vocabulary and inspected neighborhood.