For a fairly accessible description of this work see [1]. As I understand it, they're addressing the problem of the internal functioning of DNNs being hard to i

For a fairly accessible description of this work see [1]. As I understand it, they're addressing the problem of the internal functioning of DNNs being hard to interpret, by using composable NN modules. [1] http://news.mit.edu/2018/mit-lincoln-laboratory-ai-system-solves-problems-through-human-reasoning-0911 http://news.mit.edu/2018/mit-lincoln-laboratory-ai-system-so...