Masking Strategies for JEPA

What I worked on

Tried to figure out how to apply masking for JEPA training where each frame of a driving video is partially hidden. Looked at determinism, seeding, and using unfold to create patches.

What I noticed

• Masking can be applied on tensors or images
• Determinism depends on consistent seeding
• Using unfold can handle patch creation efficiently

”Aha” Moment

na

What still feels messy

Still not sure how much of the image should be masked or how patch size affects learning stability.

Next step

Test patch coverage percentages and observe reconstruction loss behavior.