Fourier power predicted famous face labels across resized ds000117 v2.0.0 picture sets with separate training and test source identities

Fourier power features predicted famous versus unfamiliar face labels in OpenNeuro ds000117 version 2.0.0. We used 299 source identity codes after removing the duplicate identity reported by the authors. A fixed ridge classifier used 36 power features from pictures resized to 128 by 128 pixels. Training and test identities were separate. The main test trained on M/EEG pictures and tested on fMRI pictures. Balanced accuracy was 56.91%. A test with 9,999 identity-label permutations gave p = 0.0040. The maximum-statistic p value across four planned conditions was 0.0479. Prediction within each picture set reached 62.87% and 64.20%. The source identity mapping and label exchangeability define the test assumptions. Additional checks gave 56.60% transfer accuracy for identities with fMRI event entries, and 54.54% without resizing. The adjusted p values in those checks were above 0.05. Thus, transfer evidence is modest and sensitive to processing. The event audit found 12 supplied fMRI pictures without stored event entries. The package preserves all 600 pictures, all 240 event tables, source hashes, code, features, predictions, and permutation results. A second run reproduced all 21 numerical result files. The features provide a stimulus control for neural decoding. This image result does not establish the cause of any neural effect.

face images · Fourier power · famous faces · image classification · OpenNeuro · cross-set prediction