GloVe cosine reduced recall choice loss in 68 selected ds004789 test people with a fixed word pool and ridge penalty

We tested a fixed external word measure in OpenNeuro ds004789 version 3.1.0. We fit models in 67 people and tested them in 68 other people. The reference used signed study lag, target position, and target word terms. One added GloVe cosine term reduced test loss from 1.979506 to 1.955153 nats. The gain was 0.024353 nats per scored transition, with equal list and person weights. A 1,999-draw person bootstrap refit both models. Its approximate 95% interval was [0.015450, 0.034760] nats. All five planned checks had positive point gains. The word pool, split, selection rules, and penalties were fixed. Selection used later recall records. Thus, the result applies to this retrospective conditional test. It does not show a causal memory process or a brain signal. The package preserves source records, word vectors, code, all fitted models, test probabilities, bootstrap draws, and exact repeat checks.

free recall · memory search · GloVe · word embeddings · cosine similarity · conditional choice model · cross-validation · bootstrap · OpenNeuro