PCA replay reportedly reaches 91.1% on a toy benchmark

A user describes PCA replay as summarizing each class’s pixels with a mean and 10 principal-component directions, sampling synthetic inputs, and penalizing changes to predictions saved before the next task. They report 90.2%, but note that a simultaneous switch to summed binary cross-entropy across all classes means replay’s effect isn’t isolated.

With stochastic gradient descent (SGD) and PCA replay, increasing training from two to four passes per task reportedly reaches 91.1%. Joint training gets 95.2% with 10 passes through all tasks’ data together. The user cautions that the different budgets, single toy benchmark and three seeds make this a useful reference—not a ceiling.