A statistical learning model for figural after-effects