I found that influential training digits were usually more sloppy / unclear compared to the average MNIST digit, and shared some resemblance with the query image.
It’s pbzcnevat gb gur arnerfg cbvagf ba gur obhaqnel bs qvtvg-pyhfgref! Bonus points if you made your observation without that interpretation in mind. What if you do Jacobian regularization?
It’s pbzcnevat gb gur arnerfg cbvagf ba gur obhaqnel bs qvtvg-pyhfgref! Bonus points if you made your observation without that interpretation in mind. What if you do Jacobian regularization?
How do you know?
It’s the same training datums I would look at to resolve an ambiguous case.