Extended Data Fig. 3: Mathematical intuition for the counterexample strategy, exemplified for linear classifiers. | Nature Machine Intelligence

Extended Data Fig. 3: Mathematical intuition for the counterexample strategy, exemplified for linear classifiers.

From: Making deep neural networks right for the right scientific reasons by interacting with their explanations

Extended Data Fig. 3

Two data features are shown, Ï•1 and Ï•2, of which only the first is truly relevant. a, The positive example xi is not enough to disambiguate between the red and green classifiers. b, Counterexamples xi,â„“ are obtained by randomizing the irrelevant feature while keeping the label of xi. The counterexamples approximate a (local) orthogonality constraint. c, The red classifier is inconsistent with the counterexamples and eliminated. See the Methods section Explanatory Interactive Learning with counterexamples for details. (Best viewed in colour).

Back to article page