Research Papers

Integrating Prior Knowledge in Post-hoc Explanations

Adulam Jeyasothy, Thibault Laugel, Marie-Jeanne Lesot +2

In the field of eXplainable Artificial Intelligence (XAI), post-hoc interpretability methods aim at explaining to a user the predictions of a trained decision model. Integrating prior knowledge into such interpretability methods aims at improving the explanation understandability and allowing for personalised explanations adapted to each user. In this paper, we propose to define a cost function that explicitly integrates prior knowledge into the interpretability objectives: we present a general framework for the optimization problem of post-hoc interpretability methods, and show that user knowledge can thus be integrated to any method by adding a compatibility term in the cost function. We instantiate the proposed formalization in the case of counterfactual explanations and propose a new interpretability method called Knowledge Integration in Counterfactual Explanation (KICE) to optimize it. The paper performs an experimental study on several benchmark data sets to characterize the counterfactual instances generated by KICE, as compared to reference methods.

EVOTER: Evolution of Transparent Explainable Rule-sets

Hormoz Shahrzad, Babak Hodjat, Risto Miikkulainen

Most AI systems are black boxes generating reasonable outputs for given inputs. Some domains, however, have explainability and trustworthiness requirements that cannot be directly met by these approaches. Various methods have therefore been developed to interpret black-box models after training. This paper advocates an alternative approach where the models are transparent and explainable to begin with. This approach, EVOTER, evolves rule-sets based on simple logical expressions. The approach is evaluated in several prediction/classification and prescription/policy search domains with and without a surrogate. It is shown to discover meaningful rule sets that perform similarly to black-box models. The rules can provide insight into the domain, and make biases hidden in the data explicit. It may also be possible to edit them directly to remove biases and add constraints. EVOTER thus forms a promising foundation for building trustworthy AI systems for real-world applications in the future.

Lung cancer has the highest mortality rate of deadly cancers in the world. Early detection is essential to treatment of lung cancer. However, detection and accurate diagnosis of pulmonary nodules depend heavily on the experiences of radiologists and can be a heavy workload for them. Computer-aided diagnosis (CAD) systems have been developed to assist radiologists in nodule detection and diagnosis, greatly easing the workload while increasing diagnosis accuracy. Recent development of deep learning, greatly improved the performance of CAD systems. However, lack of model reliability and interpretability remains a major obstacle for its large-scale clinical application. In this work, we proposed a multi-task explainable deep-learning model for pulmonary nodule diagnosis. Our neural model can not only predict lesion malignancy but also identify relevant manifestations. Further, the location of each manifestation can also be visualized for visual interpretability. Our proposed neural model achieved a test AUC of 0.992 on LIDC public dataset and a test AUC of 0.923 on our in-house dataset. Moreover, our experimental results proved that by incorporating manifestation identification tasks into the multi-task model, the accuracy of the malignancy classification can also be improved. This multi-task explainable model may provide a scheme for better interaction with the radiologists in a clinical environment.

Multi-modal multi-objective model-based genetic programming to find multiple diverse high-quality models

E. M. C. Sijben, T. Alderliesten, P. A. N. Bosman

Explainable artificial intelligence (XAI) is an important and rapidly expanding research topic. The goal of XAI is to gain trust in a machine learning (ML) model through clear insights into how the model arrives at its predictions. Genetic programming (GP) is often cited as being uniquely well-suited to contribute to XAI because of its capacity to learn (small) symbolic models that have the potential to be interpreted. Nevertheless, like many ML algorithms, GP typically results in a single best model. However, in practice, the best model in terms of training error may well not be the most suitable one as judged by a domain expert for various reasons, including overfitting, multiple different models existing that have similar accuracy, and unwanted errors on particular data points due to typical accuracy measures like mean squared error. Hence, to increase chances that domain experts deem a resulting model plausible, it becomes important to be able to explicitly search for multiple, diverse, high-quality models that trade-off different meanings of accuracy. In this paper, we achieve exactly this with a novel multi-modal multi-tree multi-objective GP approach that extends a modern model-based GP algorithm known as GP-GOMEA that is already effective at searching for small expressions.

Sparse Subspace Clustering for Concept Discovery (SSCCD)

Johanna Vielhaben, Stefan Blücher, Nils Strodthoff

Concepts are key building blocks of higher level human understanding. Explainable AI (XAI) methods have shown tremendous progress in recent years, however, local attribution methods do not allow to identify coherent model behavior across samples and therefore miss this essential component. In this work, we study concept-based explanations and put forward a new definition of concepts as low-dimensional subspaces of hidden feature layers. We novelly apply sparse subspace clustering to discover these concept subspaces. Moving forward, we derive insights from concept subspaces in terms of localized input (concept) maps, show how to quantify concept relevances and lastly, evaluate similarities and transferability between concepts. We empirically demonstrate the soundness of the proposed Sparse Subspace Clustering for Concept Discovery (SSCCD) method for a variety of different image classification tasks. This approach allows for deeper insights into the actual model behavior that would remain hidden from conventional input-level heatmaps.

Legislation and public sentiment throughout the world have promoted fairness metrics, explainability, and interpretability as prescriptions for the responsible development of ethical artificial intelligence systems. Despite the importance of these three pillars in the foundation of the field, they can be challenging to operationalize and attempts to solve the problems in production environments often feel Sisyphean. This difficulty stems from a number of factors: fairness metrics are computationally difficult to incorporate into training and rarely alleviate all of the harms perpetrated by these systems. Interpretability and explainability can be gamed to appear fair, may inadvertently reduce the privacy of personal information contained in training data, and increase user confidence in predictions -- even when the explanations are wrong. In this work, we propose a framework for responsibly developing artificial intelligence systems by incorporating lessons from the field of information security and the secure development lifecycle to overcome challenges associated with protecting users in adversarial settings. In particular, we propose leveraging the concepts of threat modeling, design review, penetration testing, and incident response in the context of developing AI systems as ways to resolve shortcomings in the aforementioned methods.

Deep reinforcement learning has led to numerous notable results in robotics. However, deep neural networks (DNNs) are unintuitive, which makes it difficult to understand their predictions and strongly limits their potential for real-world applications due to economic, safety, and assurance reasons. To remedy this problem, a number of explainable AI methods have been presented, such as SHAP and LIME, but these can be either be too costly to be used in real-time robotic applications or provide only local explanations. In this paper, the main contribution is the use of a linear model tree (LMT) to approximate a DNN policy, originally trained via proximal policy optimization(PPO), for an autonomous surface vehicle with five control inputs performing a docking operation. The two main benefits of the proposed approach are: a) LMTs are transparent which makes it possible to associate directly the outputs (control actions, in our case) with specific values of the input features, b) LMTs are computationally efficient and can provide information in real-time. In our simulations, the opaque DNN policy controls the vehicle and the LMT runs in parallel to provide explanations in the form of feature attributions. Our results indicate that LMTs can be a useful component within digital assurance frameworks for autonomous ships.

Rethinking the Authorship Verification Experimental Setups

Florin Brad, Andrei Manolache, Elena Burceanu +3

One of the main drivers of the recent advances in authorship verification is the PAN large-scale authorship dataset. Despite generating significant progress in the field, inconsistent performance differences between the closed and open test sets have been reported. To this end, we improve the experimental setup by proposing five new public splits over the PAN dataset, specifically designed to isolate and identify biases related to the text topic and to the author's writing style. We evaluate several BERT-like baselines on these splits, showing that such models are competitive with authorship verification state-of-the-art methods. Furthermore, using explainable AI, we find that these baselines are biased towards named entities. We show that models trained without the named entities obtain better results and generalize better when tested on DarkReddit, our new dataset for authorship verification.

This paper provides empirical concerns about post-hoc explanations of black-box ML models, one of the major trends in AI explainability (XAI), by showing its lack of interpretability and societal consequences. Using a representative consumer panel to test our assumptions, we report three main findings. First, we show that post-hoc explanations of black-box model tend to give partial and biased information on the underlying mechanism of the algorithm and can be subject to manipulation or information withholding by diverting users' attention. Secondly, we show the importance of tested behavioral indicators, in addition to self-reported perceived indicators, to provide a more comprehensive view of the dimensions of interpretability. This paper contributes to shedding new light on the actual theoretical debate between intrinsically transparent AI models and post-hoc explanations of black-box complex models-a debate which is likely to play a highly influential role in the future development and operationalization of AI systems.

Enabling humans to identify potential flaws in an agent's decision making is an important Explainable AI application. We consider identifying such flaws in a planning-based deep reinforcement learning (RL) agent for a complex real-time strategy game. In particular, the agent makes decisions via tree search using a learned model and evaluation function over interpretable states and actions. This gives the potential for humans to identify flaws at the level of reasoning steps in the tree, even if the entire reasoning process is too complex to understand. However, it is unclear whether humans will be able to identify such flaws due to the size and complexity of trees. We describe a user interface and case study, where a small group of AI experts and developers attempt to identify reasoning flaws due to inaccurate agent learning. Overall, the interface allowed the group to identify a number of significant flaws of varying types, demonstrating the promise of this approach.

The need of predictive maintenance comes with an increasing number of incidents reported by monitoring systems and equipment/software users. In the front line, on-call engineers (OCEs) have to quickly assess the degree of severity of an incident and decide which service to contact for corrective actions. To automate these decisions, several predictive models have been proposed, but the most efficient models are opaque (say, black box), strongly limiting their adoption. In this paper, we propose an efficient black box model based on 170K incidents reported to our company over the last 7 years and emphasize on the need of automating triage when incidents are massively reported on thousands of servers running our product, an ERP. Recent developments in eXplainable Artificial Intelligence (XAI) help in providing global explanations to the model, but also, and most importantly, with local explanations for each model prediction/outcome. Sadly, providing a human with an explanation for each outcome is not conceivable when dealing with an important number of daily predictions. To address this problem, we propose an original data-mining method rooted in Subgroup Discovery, a pattern mining technique with the natural ability to group objects that share similar explanations of their black box predictions and provide a description for each group. We evaluate this approach and present our preliminary results which give us good hope towards an effective OCE's adoption. We believe that this approach provides a new way to address the problem of model agnostic outcome explanation.

A central goal of explainable artificial intelligence (XAI) is to improve the trust relationship in human-AI interaction. One assumption underlying research in transparent AI systems is that explanations help to better assess predictions of machine learning (ML) models, for instance by enabling humans to identify wrong predictions more efficiently. Recent empirical evidence however shows that explanations can have the opposite effect: When presenting explanations of ML predictions humans often tend to trust ML predictions even when these are wrong. Experimental evidence suggests that this effect can be attributed to how intuitive, or human, an AI or explanation appears. This effect challenges the very goal of XAI and implies that responsible usage of transparent AI methods has to consider the ability of humans to distinguish machine generated from human explanations. Here we propose a quantitative metric for XAI methods based on Turing's imitation game, a Turing Test for Transparency. A human interrogator is asked to judge whether an explanation was generated by a human or by an XAI method. Explanations of XAI methods that can not be detected by humans above chance performance in this binary classification task are passing the test. Detecting such explanations is a requirement for assessing and calibrating the trust relationship in human-AI interaction. We present experimental results on a crowd-sourced text classification task demonstrating that even for basic ML models and XAI approaches most participants were not able to differentiate human from machine generated explanations. We discuss ethical and practical implications of our results for applications of transparent ML.