JEPA Learns What the Mask Leaves Unrecoverable
The authors study how JEPA masking geometry determines what can be recovered from context. They model a mask as a linear measurement: atoms whose support lies in the hidden region fall in the null space in a wavelet basis. JEPA must predict only sufficiency, enabling shortcuts with a moving-average target encoder; removing the shortcut depends on coarse-scale content left unrecoverable and on reachable context. They test predictions in 151 pre-training runs.
