Yearbook

// 01  the dataset

Yearbook is a collection of American high-school yearbook portraits. We use 37,921 frontal photos from to , each labelled male or female, with the two classes roughly balanced over time.

The target (the male/female label) is held fixed, so the temporal signal is dominated by covariate and concept drift: hairstyles, eyewear, framing, and photographic process drift steadily across the century, with the sharpest stylistic break around the mid-century decades. The benchmark asks whether a model trained on portraits up to one year still recognises portraits from later years.

Yearbookimages

input a high-school portrait photo

predict male or female femalemale

a single binary label

what changes how people look: hairstyles, fashion, photo quality

19052013

104 yearly slices

scored by accuracydataset source ↗

Exploratory analysis

// 02  the models

Yearbook is evaluated with 21 models: twelve trained end to end from raw pixels (four families at three sizes) and nine frozen pretrained vision encoders, each fitted with a single linear head. The drift matrices report all 21 models across every training cutoff.

MLP pixels as an unordered vector

3 models

CNN / ResNet locality and translation invariance

6 models

ViT patches compared by self-attention

3 models

Frozen features from a frozen pretrained encoder

9 models

Trained end-to-end

Frozen pretrained encoders