JEPA (I-JEPA, V-JEPA) — LeCun's alternative to generative SSL.

hard

Answer

$Predict abstract features (not pixels) of masked regions in embedding space.$
$Sidesteps pixel-level detail (irrelevant + wasteful) and models context predictively.$
$I-JEPA (images 2023): predicts feature representation of masked target block from context.$
$V-JEPA (video 2024): same for video.$
$Faster + more semantic representations than MAE.$
$LeCun's flagship 'world model' architecture.$

Check yourself — multiple choice

Same as MAE
Predict abstract features (not pixels) of masked regions in embedding space; faster + more semantic than MAE; LeCun's 'world model' vision (I-JEPA, V-JEPA)
Random
Not real

JEPA: predict features (not pixels) of masked regions; LeCun's world model.

#representation-learning#deep-learning

214 interview questions in this topic.