From Tamara Quinn
Your article on neural networks stated that “no one knows how the neural networks come up with their answers” and even that it is “impossible” to know how they do it (11 July, p 20). Could this be because research has, so far, focused on constructing neural networks that produce correct answers, rather than networks that also tell us how they work?
Surely it is possible to add self-monitoring and recording functions, so that the networks can reveal the paths taken to reach the observed outcome. Or is there some fundamental feature of neural networks that means this wouldn’t reveal their secrets?
London, UK
