VISReg

VISReg: Variance-Invariance-Sketching Regularization for JEPA training

Haiyu Wu, Randall Balestriero, Morgan Levine · arXiv 2026

Read paper · arXiv:2606.02572

Separate representation scale and distribution shape with sliced Gaussian matching.

What you will implement

Scale, sliced shape, centering, and multi-view invariance losses.

• 56. VISReg Scale and Distribution Shape

  Separate feature scale, distribution shape and view invariance with the correct gradient boundaries.

  Implement the paper regularizer and global-view-mean invariance with supplied directions and features; encoders and image pretraining are outside this module.