On multi-view feature learning