From Alan Bundy
Edinburgh, UK
Brian Horton suggests, possibly in jest, that we find out how people distinguish heroes from villains by letting an artificial intelligence watch thousands of films (Letters, 22 October). But statistical machine learning programs are notoriously unable to explain what they have learned.
If their training is successful, they will work out how to distinguish heroes and villains in previously unseen films, but will not be able to explain how they do it. This is a serious failing for some potential applications. For instance, doctors take professional responsibility for their diagnoses and treatments. They cannot accept the output of a black box without some kind of explanation.
