Closing the Gap Between Performance and Interpretability in Visual Reasoning

[andyjohnson0]

https://arxiv.org/abs/1803.05268

[andyjohnson0]

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-so...