Despite their considerable potential, concept-based explainability methods have received relatively little attention, and explaining what’s driving models’ decisions and where it’s located in the input is still an open problem. To tackle this, we revisit unsupervised concept extraction techniques for explaining the decisions of deep neural networks and present CRAFT – a framework to generate concept-based explanations for understanding individual predictions and the model’s high-level logic for whole classes. CRAFT takes advantage of a novel method for recursively decomposing higher-level concepts into more elementary ones, combined with a novel approach for better estimating the importance of identified concepts with Sobol indices. Furthermore, we show how implicit differentiation can be used to generate concept-wise attribution explanations for individual images. We further demonstrate through fidelity metrics that our proposed concept importance estimation technique is more faithful to the model than previous methods, and, through human psychophysic experiments, we confirm that our recursive decomposition can generate meaningful and accurate concepts. Finally, we illustrate CRAFT’s potential to enable the understanding of predictions of trained models on multiple use-cases by producing meaningful concept-based explanations.

The many successes of deep neural networks (DNNs) over the past decade have largely been driven by computational scale rather than insights from biological intelligence. Here, we explore if these trends have also carried concomitant improvements in explaining the visual strategies humans rely on for object recognition. We do this by comparing two related but distinct properties of visual strategies in humans and DNNs: where they believe important visual features are in images and how they use those features to categorize objects. Across 84 different DNNs trained on ImageNet and three independent datasets measuring the where and the how of human visual strategies for object recognition on those images, we find a systematic trade-off between DNN categorization accuracy and alignment with human visual strategies for object recognition. State-of-the-art DNNs are progressively becoming less aligned with humans as their accuracy improves. We rectify this growing issue with our harmonization loss: a general-purpose training routine that both aligns DNN and human visual strategies and improves categorization accuracy. Our work represents the first demonstration that the scaling laws that are guiding the design of DNNs today have also produced worse models of human vision.

In this work, we conducted psychophysics experiments at scale to evaluate the ability of human participants to leverage representative attribution methods for understanding the behavior of different image classifiers representing three real-world scenarios: (1) identifying bias, (2) characterizing novel strategy and (3) understanding failure cases. Our results demonstrate that the degree to which individual attribution methods help human participants better understand an AI system varied widely across these scenarios. This suggests a critical need for the field to move past quantitative improvements of current attribution methods.

This paper could be seen as a follow-up to Look at the variance and further improves the efficiency of black-box methods using Hilbert-Schmidt Independence Criterion (HSIC). HSIC measures the dependence between regions of an input image and the output of a model based on kernel embeddings of distributions. Our experiments show that HSIC is up to 8 times faster than the previous best black-box attribution methods while being as faithful. Importantly, we show that these advances can be transposed to efficiently and faithfully explain object detection models such as YOLOv4.

In this work, we propose the first explainability method based on formal method and therefore able to meet certification requirements. Specifically, we introduce EVA (Explaining using Verified perturbation Analysis) -- the first explainability method guarantee to have an exhaustive exploration of a perturbation space. We leverage the beneficial properties of verified perturbation analysis -- time efficiency, tractability and guaranteed complete coverage of a manifold -- to efficiently characterize the input variables that are most likely to drive the model decision. We evaluate the approach systematically and demonstrate state-of-the-art results on multiple benchmarks.

We argue that, when learning a 1-Lipschitz neural network with the dual loss of an optimal transportation problem, the gradient of the model is both the direction of the transportation plan and the direction to the closest adversarial attack. Traveling along the gradient to the decision boundary is no more an adversarial attack but becomes a counterfactual explanation, explicitly transporting from one class to the other. Through extensive experiments on XAI metrics, we find that the simple saliency map method, applied on such networks, becomes a reliable explanation! The proposed networks were already known to be certifiably robust, and we prove that they are also tailored for explainability.

For more than a year, I worked alongside my thesis to implement more than fifty explainability papers. Xplique is the result of this work, it is a library that I develop and maintain. The library is composed of several modules: (1) the Attributions Methods module, (2) The Feature Visualization module, (3) The Concepts module and (4) the Metrics module.

We describe a novel and efficient black-box attribution method which is grounded in Sensitivity Analysis and uses Sobol indices. Beyond modeling the individual contributions of image regions, Sobol indices provide an efficient way to capture higher-order interactions between image regions and their contributions to a neural network's prediction through the lens of variance.

We propose two new measures to evaluate explanations borrowed from the field of algorithmic stability: mean generalizability MeGe and relative consistency ReCo. We conduct extensive experiments on different network architectures, common explainability methods, and several image datasets to demonstrate the benefits of the proposed metrics and show that popular fidelity measures are not sufficient to guarantee trustworthy explanations. Finally, we found that 1-Lipschitz networks produce explanations with higher MeGe and ReCo than common neural networks.