PCA replay's reported 90.2% result comes with a loss-function caveat

A user reports reaching 90.2% with generative PCA replay: summarize each class's pixels with a mean and 10 principal component directions, sample synthetic inputs, then penalize changes to predictions saved before the next task. But the setup also switched to summed, all-class binary cross-entropy (BCE) loss, so the user cautions that replay's effect wasn't isolated.

The same thread illustrates why loss scaling matters: the user reports that summing BCE across two outputs instead of averaging it dropped accuracy from 58.1% to 51.2%. At a fixed stochastic gradient descent learning rate, they explain, that change doubles the update.