[MON 25 DEC 17] UNDERSTANDING AI (5)

Marco Ribeiro, a graduate student at the University of Washington in Seattle, probes DNNs using what are called "counterfactual probes" -- a trick related to that used by the MIT researchers to spoof DNNs. The idea is to give the DNN inputs carefully varied over a wide range, and see which way the outputs of the DNN jump. For example, consider a neural network that reads in the text of movie reviews, and then flags those that give a movie a thumb's-up. To do this, the DNN would first be trained by being given reviews flagged as being positive, plus reviews flagged as being negative, with the DNN's ability then using the training from these examples to flag reviews appropriately itself.