Digit-learning model reportedly scores 19.4% across ten digits, 97.5% within pairs

A user describes training a small neural network on handwritten-digit pairs in sequence: 0/1, 2/3, 4/5, 6/7 and 8/9. Each stage used only that pair’s real training images, followed by a final test on all ten digits.

With cross-entropy over all ten outputs and SGD, the user reports 19.4% final accuracy—but 97.5% when answers were restricted to the correct digit pair. With Adam, the corresponding results were 19.6% and 71.3%. Results averaged three seeds, with the setup starting at two epochs per task.

The user’s takeaway: poor ten-class accuracy can hide retained within-pair discrimination.