An interesting comment after the original article — training a network of this type on sudoku solutions lead to the network having false memories of correct solutions when presented with a new partially solved sudoku.