Can a model recover
what it learned to ignore?
Say a model learns that color has nothing to do with motion. Later, color starts predicting friction. Can the model use that information again?
I’m looking at this in models that learn representations through prediction. First, I need to tell whether a feature disappeared from the representation or the predictor just stopped using it. Those are different problems.
I want to test whether a model can use that information again while keeping its earlier predictions useful, with the same memory and compute budget.