Learning Curves and Gradient Behavior

Hello, thanks you for your work.
I have tried to run JEPA as is, with the same configuration, but using 2K samples from Kinetics dataset as a starting point. Subsequently, I plotted the losses (jepa and regression, although regression coeff was set to 0) along the global encoder and predictor norms. Below are the plots.


Are such plots familiar or normal?

Reproduce JEPA with the stated configuration and 2K Kinetics samples, then inspect the plotted JEPA and regression losses alongside global encoder and predictor norms. No source file or test is named; the issue is done when the observed curves are classified as expected or a specific training problem and follow-up change are identified.

Written by the indexing model from the issue text.