Research Papers

This paper follows calls for critical approaches to computing and conceptualisations of intersectional, feminist, decolonial HCI and AI design and asks what a feminist intersectional perspective in HCXAI research and design might look like. Sketching out initial research directions and implications for explainable AI design, it suggests that explainability from a feminist perspective would include the fostering of response-ability - the capacity to critically evaluate and respond to AI systems - and would centre marginalised perspectives.

Revealing Similar Semantics Inside CNNs: An Interpretable Concept-based Comparison of Feature Spaces

Georgii Mikriukov, Gesina Schwalbe, Christian Hellert +1

Safety-critical applications require transparency in artificial intelligence (AI) components, but widely used convolutional neural networks (CNNs) widely used for perception tasks lack inherent interpretability. Hence, insights into what CNNs have learned are primarily based on performance metrics, because these allow, e.g., for cross-architecture CNN comparison. However, these neglect how knowledge is stored inside. To tackle this yet unsolved problem, our work proposes two methods for estimating the layer-wise similarity between semantic information inside CNN latent spaces. These allow insights into both the flow and likeness of semantic information within CNN layers, and into the degree of their similarity between different network architectures. As a basis, we use two renowned explainable artificial intelligence (XAI) techniques, which are used to obtain concept activation vectors, i.e., global vector representations in the latent space. These are compared with respect to their activation on test inputs. When applied to three diverse object detectors and two datasets, our methods reveal that (1) similar semantic concepts are learned regardless of the CNN architecture, and (2) similar concepts emerge in similar relative layer depth, independent of the total number of layers. Finally, our approach poses a promising step towards semantic model comparability and comprehension of how different CNNs process semantic information.

As the use of deep neural networks continues to grow, understanding their behaviour has become more crucial than ever. Post-hoc explainability methods are a potential solution, but their reliability is being called into question. Our research investigates the response of post-hoc visual explanations to naturally occurring transformations, often referred to as augmentations. We anticipate explanations to be invariant under certain transformations, such as changes to the colour map while responding in an equivariant manner to transformations like translation, object scaling, and rotation. We have found remarkable differences in robustness depending on the type of transformation, with some explainability methods (such as LRP composites and Guided Backprop) being more stable than others. We also explore the role of training with data augmentation. We provide evidence that explanations are typically less robust to augmentation than classification performance, regardless of whether data augmentation is used in training or not.

Explainable AI (XAI) methods provide explanations of AI models, but our understanding of how they compare with human explanations remains limited. In image classification, we found that humans adopted more explorative attention strategies for explanation than the classification task itself. Two representative explanation strategies were identified through clustering: One involved focused visual scanning on foreground objects with more conceptual explanations diagnostic for inferring class labels, whereas the other involved explorative scanning with more visual explanations rated higher for effectiveness. Interestingly, XAI saliency-map explanations had the highest similarity to the explorative attention strategy in humans, and explanations highlighting discriminative features from invoking observable causality through perturbation had higher similarity to human strategies than those highlighting internal features associated with higher class score. Thus, humans differ in information and strategy use for explanations, and XAI methods that highlight features informing observable causality match better with human explanations, potentially more accessible to users.

Trust is a key motivation in developing explainable artificial intelligence (XAI). However, researchers attempting to measure trust in AI face numerous challenges, such as different trust conceptualizations, simplified experimental tasks that may not induce uncertainty as a prerequisite for trust, and the lack of validated trust questionnaires in the context of AI. While acknowledging these issues, we have identified a further challenge that currently seems underappreciated - the potential distinction between trust as one construct and \emph{distrust} as a second construct independent of trust. While there has been long-standing academic discourse for this distinction and arguments for both the one-dimensional and two-dimensional conceptualization of trust, distrust seems relatively understudied in XAI. In this position paper, we not only highlight the theoretical arguments for distrust as a distinct construct from trust but also contextualize psychometric evidence that likewise favors a distinction between trust and distrust. It remains to be investigated whether the available psychometric evidence is sufficient for the existence of distrust or whether distrust is merely a measurement artifact. Nevertheless, the XAI community should remain receptive to considering trust and distrust for a more comprehensive understanding of these two relevant constructs in XAI.

The AI Act proposal: a new right to technical interpretability?

Chiara Gallese

The debate about the concept of the so called right to explanation in AI is the subject of a wealth of literature. It has focused, in the legal scholarship, on art. 22 GDPR and, in the technical scholarship, on techniques that help explain the output of a certain model (XAI). The purpose of this work is to investigate if the new provisions introduced by the proposal for a Regulation laying down harmonised rules on artificial intelligence (AI Act), in combination with Convention 108 plus and GDPR, are enough to indicate the existence of a right to technical explainability in the EU legal framework and, if not, whether the EU should include it in its current legislation. This is a preliminary work submitted to the online event organised by the Information Society Law Center and it will be later developed into a full paper.

Developing automated and smart software vulnerability detection models has been receiving great attention from both research and development communities. One of the biggest challenges in this area is the lack of code samples for all different programming languages. In this study, we address this issue by proposing a transfer learning technique to leverage available datasets and generate a model to detect common vulnerabilities in different programming languages. We use C source code samples to train a Convolutional Neural Network (CNN) model, then, we use Java source code samples to adopt and evaluate the learned model. We use code samples from two benchmark datasets: NIST Software Assurance Reference Dataset (SARD) and Draper VDISC dataset. The results show that proposed model detects vulnerabilities in both C and Java codes with average recall of 72\%. Additionally, we employ explainable AI to investigate how much each feature contributes to the knowledge transfer mechanisms between C and Java in the proposed model.

Can machines think? Since Alan Turing asked this question in 1950, nobody is able to give a direct answer, due to the lack of solid mathematical foundations for general intelligence. In this paper, we introduce a categorical framework towards this goal, with two main results. First, we investigate object representation through presheaves, introducing the notion of self-state awareness as a categorical analogue to self-consciousness, along with corresponding algorithms for its enforcement and evaluation. Secondly, we extend object representation to scenario representation using diagrams and limits, which then become building blocks for mathematical modeling, interpretability and AI safety. As an ancillary result, our framework introduces various categorical invariance properties that can serve as the alignment signals for model training.

Explaining artificial intelligence or machine learning models is increasingly important. To use such data-driven systems wisely we must understand how they interact with the world, including how they depend causally on data inputs. In this work we develop Causal Dependence Plots (CDPs) to visualize how one variable--an outcome--depends on changes in another variable--a predictor--$\textit{along with any consequent causal changes in other predictor variables}$. Crucially, CDPs differ from standard methods based on holding other predictors constant or assuming they are independent. CDPs make use of an auxiliary causal model because causal conclusions require causal assumptions. With simulations and real data experiments, we show CDPs can be combined in a modular way with methods for causal learning or sensitivity analysis. Since people often think causally about input-output dependence, CDPs can be powerful tools in the xAI or interpretable machine learning toolkit and contribute to applications like scientific machine learning and algorithmic fairness.

Finding important features that contribute to the prediction of neural models is an active area of research in explainable AI. Neural models are opaque and finding such features sheds light on a better understanding of their predictions. In contrast, in this work, we present an inverse perspective of distractor features: features that cast doubt about the prediction by affecting the model's confidence in its prediction. Understanding distractors provide a complementary view of the features' relevance in the predictions of neural models. In this paper, we apply a reduction-based technique to find distractors and provide our preliminary results of their impacts and types. Our experiments across various tasks, models, and datasets of code reveal that the removal of tokens can have a significant impact on the confidence of models in their predictions and the categories of tokens can also play a vital role in the model's confidence. Our study aims to enhance the transparency of models by emphasizing those tokens that significantly influence the confidence of the models.

Towards Modelling and Verification of Social Explainable AI

Damian Kurpiewski, Wojciech Jamroga, Teofil Sidoruk

Social Explainable AI (SAI) is a new direction in artificial intelligence that emphasises decentralisation, transparency, social context, and focus on the human users. SAI research is still at an early stage. Consequently, it concentrates on delivering the intended functionalities, but largely ignores the possibility of unwelcome behaviours due to malicious or erroneous activity. We propose that, in order to capture the breadth of relevant aspects, one can use models and logics of strategic ability, that have been developed in multi-agent systems. Using the STV model checker, we take the first step towards the formal modelling and verification of SAI environments, in particular of their resistance to various types of attacks by compromised AI modules.

The recent spike in certified Artificial Intelligence (AI) tools for healthcare has renewed the debate around adoption of this technology. One thread of such debate concerns Explainable AI (XAI) and its promise to render AI devices more transparent and trustworthy. A few voices active in the medical AI space have expressed concerns on the reliability of Explainable AI techniques and especially feature attribution methods, questioning their use and inclusion in guidelines and standards. Despite valid concerns, we argue that existing criticism on the viability of post-hoc local explainability methods throws away the baby with the bathwater by generalizing a problem that is specific to image data. We begin by characterizing the problem as a lack of semantic match between explanations and human understanding. To understand when feature importance can be used reliably, we introduce a distinction between feature importance of low- and high-level features. We argue that for data types where low-level features come endowed with a clear semantics, such as tabular data like Electronic Health Records (EHRs), semantic match can be obtained, and thus feature attribution methods can still be employed in a meaningful and useful way. Finally, we sketch a procedure to test whether semantic match has been achieved.