Research Papers

DeDUCE: Generating Counterfactual Explanations Efficiently

Benedikt Höltgen, Lisa Schut, Jan M. Brauner +1

When an image classifier outputs a wrong class label, it can be helpful to see what changes in the image would lead to a correct classification. This is the aim of algorithms generating counterfactual explanations. However, there is no easily scalable method to generate such counterfactuals. We develop a new algorithm providing counterfactual explanations for large image classifiers trained with spectral normalisation at low computational cost. We empirically compare this algorithm against baselines from the literature; our novel algorithm consistently finds counterfactuals that are much closer to the original inputs. At the same time, the realism of these counterfactuals is comparable to the baselines. The code for all experiments is available at https://github.com/benedikthoeltgen/DeDUCE.

Expose Uncertainty, Instill Distrust, Avoid Explanations: Towards Ethical Guidelines for AI

Claudio S. Pinhanez

In this position paper, I argue that the best way to help and protect humans using AI technology is to make them aware of the intrinsic limitations and problems of AI algorithms. To accomplish this, I suggest three ethical guidelines to be used in the presentation of results, mandating AI systems to expose uncertainty, to instill distrust, and, contrary to traditional views, to avoid explanations. The paper does a preliminary discussion of the guidelines and provides some arguments for their adoption, aiming to start a debate in the community about AI ethics in practice.

To explain the decision of any model, we extend the notion of probabilistic Sufficient Explanations (P-SE). For each instance, this approach selects the minimal subset of features that is sufficient to yield the same prediction with high probability, while removing other features. The crux of P-SE is to compute the conditional probability of maintaining the same prediction. Therefore, we introduce an accurate and fast estimator of this probability via random Forests for any data $(\boldsymbol{X}, Y)$ and show its efficiency through a theoretical analysis of its consistency. As a consequence, we extend the P-SE to regression problems. In addition, we deal with non-discrete features, without learning the distribution of $\boldsymbol{X}$ nor having the model for making predictions. Finally, we introduce local rule-based explanations for regression/classification based on the P-SE and compare our approaches w.r.t other explainable AI methods. These methods are available as a Python package at \url{www.github.com/salimamoukou/acv00}.

We propose a novel approach to explainable AI (XAI) based on the concept of "instruction" from neural networks. In this case study, we demonstrate how a superhuman neural network might instruct human trainees as an alternative to traditional approaches to XAI. Specifically, an AI examines human actions and calculates variations on the human strategy that lead to better performance. Experiments with a JHU/APL-developed AI player for the cooperative card game Hanabi suggest this technique makes unique contributions to explainability while improving human performance. One area of focus for Instructive AI is in the significant discrepancies that can arise between a human's actual strategy and the strategy they profess to use. This inaccurate self-assessment presents a barrier for XAI, since explanations of an AI's strategy may not be properly understood or implemented by human recipients. We have developed and are testing a novel, Instructive AI approach that estimates human strategy by observing human actions. With neural networks, this allows a direct calculation of the changes in weights needed to improve the human strategy to better emulate a more successful AI. Subjected to constraints (e.g. sparsity) these weight changes can be interpreted as recommended changes to human strategy (e.g. "value A more, and value B less"). Instruction from AI such as this functions both to help humans perform better at tasks, but also to better understand, anticipate, and correct the actions of an AI. Results will be presented on AI instruction's ability to improve human decision-making and human-AI teaming in Hanabi.

In this extended abstract, we will present and discuss opportunities and challenges brought about by a new deep learning method by AUC maximization (aka \underline{\bf D}eep \underline{\bf A}UC \underline{\bf M}aximization or {\bf DAM}) for medical image classification. Since AUC (aka area under ROC curve) is a standard performance measure for medical image classification, hence directly optimizing AUC could achieve a better performance for learning a deep neural network than minimizing a traditional loss function (e.g., cross-entropy loss). Recently, there emerges a trend of using deep AUC maximization for large-scale medical image classification. In this paper, we will discuss these recent results by highlighting (i) the advancements brought by stochastic non-convex optimization algorithms for DAM; (ii) the promising results on various medical image classification problems. Then, we will discuss challenges and opportunities of DAM for medical image classification from three perspectives, feature learning, large-scale optimization, and learning trustworthy AI models.

Facial expression recognition (FER) is a topic attracting significant research in both psychology and machine learning with a wide range of applications. Despite a wealth of research on human FER and considerable progress in computational FER made possible by deep neural networks (DNNs), comparatively less work has been done on comparing the degree to which DNNs may be comparable to human performance. In this work, we compared the recognition performance and attention patterns of humans and machines during a two-alternative forced-choice FER task. Human attention was here gathered through click data that progressively uncovered a face, whereas model attention was obtained using three different popular techniques from explainable AI: CAM, GradCAM and Extremal Perturbation. In both cases, performance was gathered as percent correct. For this task, we found that humans outperformed machines quite significantly. In terms of attention patterns, we found that Extremal Perturbation had the best overall fit with the human attention map during the task.

In this work we use the AUCMEDI-Framework to train a deep neural network to classify chest X-ray images as either normal or viral pneumonia. Stratified k-fold cross-validation with k=3 is used to generate the validation-set and 15% of the data are set aside for the evaluation of the models of the different folds and ensembles each. A random-forest ensemble as well as a Soft-Majority-Vote ensemble are built from the predictions of the different folds. Evaluation metrics (Classification-Report, macro f1-scores, Confusion-Matrices, ROC-Curves) of the individual folds and the ensembles show that the classifier works well. Finally Grad-CAM and LIME explainable artificial intelligence (XAI) algorithms are applied to visualize the image features that are most important for the prediction. For Grad-CAM the heatmaps of the three folds are furthermore averaged for all images in order to calculate a mean XAI-heatmap. As the heatmaps of the different folds for most images differ only slightly this averaging procedure works well. However, only medical professionals can evaluate the quality of the features marked by the XAI. A comparison of the evaluation metrics with metrics of standard procedures such as PCR would also be important. Further limitations are discussed.

Extending LIME for Business Process Automation

Sohini Upadhyay, Vatche Isahagian, Vinod Muthusamy +1

AI business process applications automate high-stakes business decisions where there is an increasing demand to justify or explain the rationale behind algorithmic decisions. Business process applications have ordering or constraints on tasks and feature values that cause lightweight, model-agnostic, existing explanation methods like LIME to fail. In response, we propose a local explanation framework extending LIME for explaining AI business process applications. Empirical evaluation of our extension underscores the advantage of our approach in the business process setting.

Explainable artificial intelligence (XAI) is an emerging new domain in which a set of processes and tools allow humans to better comprehend the decisions generated by black box models. However, most of the available XAI tools are often limited to simple explanations mainly quantifying the impact of individual features to the models' output. Therefore, human users are not able to understand how the features are related to each other to make predictions, whereas the inner workings of the trained models remain hidden. This paper contributes to the development of a novel graphical explainability tool that not only indicates the significant features of the model but also reveals the conditional relationships between features and the inference capturing both the direct and indirect impact of features to the models' decision. The proposed XAI methodology, termed as gLIME, provides graphical model-agnostic explanations either at the global (for the entire dataset) or the local scale (for specific data points). It relies on a combination of local interpretable model-agnostic explanations (LIME) with graphical least absolute shrinkage and selection operator (GLASSO) producing undirected Gaussian graphical models. Regularization is adopted to shrink small partial correlation coefficients to zero providing sparser and more interpretable graphical explanations. Two well-known classification datasets (BIOPSY and OAI) were selected to confirm the superiority of gLIME over LIME in terms of both robustness and consistency over multiple permutations. Specifically, gLIME accomplished increased stability over the two datasets with respect to features' importance (76%-96% compared to 52%-77% using LIME). gLIME demonstrates a unique potential to extend the functionality of the current state-of-the-art in XAI by providing informative graphically given explanations that could unlock black boxes.

There is mounting public concern over the influence that AI based systems has in our society. Coalitions in all sectors are acting worldwide to resist hamful applications of AI. From indigenous people addressing the lack of reliable data, to smart city stakeholders, to students protesting the academic relationships with sex trafficker and MIT donor Jeffery Epstein, the questionable ethics and values of those heavily investing in and profiting from AI are under global scrutiny. There are biased, wrongful, and disturbing assumptions embedded in AI algorithms that could get locked in without intervention. Our best human judgment is needed to contain AI's harmful impact. Perhaps one of the greatest contributions of AI will be to make us ultimately understand how important human wisdom truly is in life on earth.

This paper proposes a comprehensive analysis of existing concepts coming from different disciplines tackling the notion of intelligence, namely psychology and engineering, and from disciplines aiming to regulate AI innovations, namely AI ethics and law. The aim is to identify shared notions or discrepancies to consider for qualifying AI systems. Relevant concepts are integrated into a matrix intended to help defining more precisely when and how computing tools (programs or devices) may be qualified as AI while highlighting critical features to serve a specific technical, ethical and legal assessment of challenges in AI development. Some adaptations of existing notions of AI characteristics are proposed. The matrix is a risk-based conceptual model designed to allow an empirical, flexible and scalable qualification of AI technologies in the perspective of benefit-risk assessment practices, technological monitoring and regulatory compliance: it offers a structured reflection tool for stakeholders in AI development that are engaged in responsible research and innovation.Pre-print version (achieved on May 2020)

The paper proposes a novel architecture for explainable AI based on semantic technologies and AI. We tailor the architecture for the domain of demand forecasting and validate it on a real-world case study. The provided explanations combine concepts describing features relevant to a particular forecast, related media events, and metadata regarding external datasets of interest. The knowledge graph provides concepts that convey feature information at a higher abstraction level. By using them, explanations do not expose sensitive details regarding the demand forecasting models. The explanations also emphasize actionable dimensions where suitable. We link domain knowledge, forecasted values, and forecast explanations in a Knowledge Graph. The ontology and dataset we developed for this use case are publicly available for further research.