Constrained non-negative networks for a more explainable and interpretable classification

Valentine Wargnier Dauchelle, Thomas Grenier, Françoise Durand-Dubief, François Cotton, Michaël Sdika

Keywords: Monotonic network, Constrained learning, Interpretability, Anomaly detection

Abstract: Interpretability and explainability of deep networks are essential for medical image analysis. Easily explainable networks with intrinsic properties and decisions based on radiological signs and not spurious confounders are highly desirable. The guaranteed monotonic relation between the input and the output of monotonic networks could be used to design such intrinsically explainable networks, but they are rarely used for images: state-of-the-art architectures are often very shallow due to convergence problems. Identifying the critical importance of weights initialization, we propose a recipe to transform any architecture into a trainable monotonic network. By using the monotonic property, adding a calibration and constraining the training in an unsupervised way, we propose a network more explainable with human-readable counterfactual examples but also more interpretable with a decision more based on the radiological signs of the pathology. Especially, we outperform state-of-the-art methods for weakly supervised anomaly detection.

Submission Number: 32