but that's not what the challenge is. it's more likely the database of training data is wrong. we've seen that to be very true from posts here on HN. some of the data was so horribly mis-identified, it's a wonder it works at all