Temporal ensembling

Identity Notes & Concept Scope This is the prediction-ensemble algorithm in Laine and Aila, distinguished from the same paper’s Pi-model. Its dataset-indexed stored targets and epoch-based refresh differ from Mean Teacher’s per-step weight averaging. The name does not imply an ensemble of separately trained classifiers at inference. This record describes a training algorithm, not a trained checkpoint or fixed network architecture. The inspected AIO SemiSupervisedLLM class specifically combines LLM pretraining with supervised fine-tuning; its related synonym self-training is not an exact identity for this method. AIO SelfsupervisedLearning excludes human-annotated labels, Distillation describes large-to-small compression, and RegularizationLayer is a layer: none supplies an equivalent class for this algorithm. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 230 preceding native IDs/names/aliases. Full-checkout fixed-string searches included ignored and hidden files in both worktree and main checkout; matches concern prior discovery leads and the Pi-model normalization regression. Ten exact AIO CLASS nodes, their scoped synonyms and outgoing edges, plus five neighboring native records were inspected. The corrected Unicode-aware normalizer preserves Π; the previous Pi-model/Model collision was a lexical defect, not semantic identity (review issue #58).