Abstract: Methods, systems and media for explaining a basis for a binary inference task performed by a neural network. Binary relationships are identified between features detected by the trained neural network and data samples in which those features are detected. Sets of features providing an inference context are then identified in tandem with sets of features providing the inference basis within a given inference context. Each context, basis pair potentially provides an explanation for the inference behavior of the neural network. Human perceptible representations of such pairs are generated to explain the inference behavior of the neural network, including automatically identifying similar samples sharing both the context and basis, and automatically identifying likely alternative samples sharing the context but not the basis.

Abstract: Methods, systems and media for deep neural network interpretation via rule extraction. The interpretation of the deep neural network is based on extracting one or more rules approximating classification behavior of the network. Rules are defined by identifying a set of hyperplanes through the data space that collectively define a convex polytope that separates a target class of input samples from input samples of different classes. Each rule corresponds to a set of decision boundaries between two different decision outcomes. Human-understandable representations of rules may be generated. One or more rules may be used to generate a classifier. The representations and interpretations exhibit faithfulness, robustness, and comprehensiveness relative to other known approaches.

Abstract: Methods, systems and media for deep neural network interpretation via rule extraction. The interpretation of the deep neural network is based on extracting one or more rules approximating classification behavior of the network. Rules are defined by identifying a set of hyperplanes through the data space that collectively define a convex polytope that separates a target class of input samples from input samples of different classes. Each rule corresponds to a set of decision boundaries between two different decision outcomes. Human-understandable representations of rules may be generated. One or more rules may be used to generate a classifier. The representations and interpretations exhibit faithfulness, robustness, and comprehensiveness relative to other known approaches.