Interview: 'you have unlabeled images — which SSL method should you use?'

hard

Answer

$Ask: (1) Compute budget?$
$MAE is fastest to train.$
$Contrastive (SimCLR) needs huge batches / MoCo needs momentum.$
$(2) Downstream task type?$
$Classification \to most methods work; segmentation \to DINO or MAE better (spatial features); retrieval \to contrastive (SimCLR, CLIP if paired text).$
$(3) Data volume?$
$Small \to use pretrained + fine-tune, don't retrain SSL.$
$Large \to DINOv2 or MAE.$
$(4) Domain?$
$Medical / satellite / etc — start from ImageNet-pretrained + continue SSL.$

Check yourself — multiple choice

Same for all
Depends: compute (MAE fastest) + task (DINO/MAE for seg, contrastive for retrieval) + data volume (small → fine-tune pretrained; large → DINOv2/MAE) + domain (continue SSL from ImageNet)
Random
Only SimCLR

Choose SSL: compute + task + data volume + domain fit; MAE / DINO / contrastive.

#interview#representation-learning

214 interview questions in this topic.