A model's ten-class accuracy reportedly rises from 19.4% to 58.1% after a training change

For a model trained with SGD and cross-entropy over all ten outputs, a user reports 19.4% final accuracy—or 97.5% when choices are restricted to the correct pair. With Adam, the reported figures are 19.6% and 71.3%. Keeping SGD but applying binary cross-entropy only to the current pair's two outputs raises final ten-class accuracy to 58.1%, the user says. They note that predictions can still change without direct updates to old output weights because the shared network changes.