How does this compare to just reducing the likelihood of negative samples?