Research Papers

Responsible AI design is increasingly seen as an imperative by both AI developers and AI compliance experts. One of the key tasks is envisioning AI technology uses and risks. Recent studies on the model and data cards reveal that AI practitioners struggle with this task due to its inherently challenging nature. Here, we demonstrate that leveraging a Large Language Model (LLM) can support AI practitioners in this task by enabling reflexivity, brainstorming, and deliberation, especially in the early design stages of the AI development process. We developed an LLM framework, ExploreGen, which generates realistic and varied uses of AI technology, including those overlooked by research, and classifies their risk level based on the EU AI Act regulation. We evaluated our framework using the case of Facial Recognition and Analysis technology in nine user studies with 25 AI practitioners. Our findings show that ExploreGen is helpful to both developers and compliance experts. They rated the uses as realistic and their risk classification as accurate (94.5%). Moreover, while unfamiliar with many of the uses, they rated them as having high adoption potential and transformational impact.

Multiple instance learning (MIL) is an effective and widely used approach for weakly supervised machine learning. In histopathology, MIL models have achieved remarkable success in tasks like tumor detection, biomarker prediction, and outcome prognostication. However, MIL explanation methods are still lagging behind, as they are limited to small bag sizes or disregard instance interactions. We revisit MIL through the lens of explainable AI (XAI) and introduce xMIL, a refined framework with more general assumptions. We demonstrate how to obtain improved MIL explanations using layer-wise relevance propagation (LRP) and conduct extensive evaluation experiments on three toy settings and four real-world histopathology datasets. Our approach consistently outperforms previous explanation attempts with particularly improved faithfulness scores on challenging biomarker prediction tasks. Finally, we showcase how xMIL explanations enable pathologists to extract insights from MIL models, representing a significant advance for knowledge discovery and model debugging in digital histopathology. Codes are available at: https://github.com/bifold-pathomics/xMIL.

Unveiling Hidden Factors: Explainable AI for Feature Boosting in Speech Emotion Recognition

Alaa Nfissi, Wassim Bouachir, Nizar Bouguila +1

Speech emotion recognition (SER) has gained significant attention due to its several application fields, such as mental health, education, and human-computer interaction. However, the accuracy of SER systems is hindered by high-dimensional feature sets that may contain irrelevant and redundant information. To overcome this challenge, this study proposes an iterative feature boosting approach for SER that emphasizes feature relevance and explainability to enhance machine learning model performance. Our approach involves meticulous feature selection and analysis to build efficient SER systems. In addressing our main problem through model explainability, we employ a feature evaluation loop with Shapley values to iteratively refine feature sets. This process strikes a balance between model performance and transparency, which enables a comprehensive understanding of the model's predictions. The proposed approach offers several advantages, including the identification and removal of irrelevant and redundant features, leading to a more effective model. Additionally, it promotes explainability, facilitating comprehension of the model's predictions and the identification of crucial features for emotion determination. The effectiveness of the proposed method is validated on the SER benchmarks of the Toronto emotional speech set (TESS), Berlin Database of Emotional Speech (EMO-DB), Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS), and Surrey Audio-Visual Expressed Emotion (SAVEE) datasets, outperforming state-of-the-art methods. To the best of our knowledge, this is the first work to incorporate model explainability into an SER framework. The source code of this paper is publicly available via this https://github.com/alaaNfissi/Unveiling-Hidden-Factors-Explainable-AI-for-Feature-Boosting-in-Speech-Emotion-Recognition.

AI Risk Management Should Incorporate Both Safety and Security

Xiangyu Qi, Yangsibo Huang, Yi Zeng +22

The exposure of security vulnerabilities in safety-aligned language models, e.g., susceptibility to adversarial attacks, has shed light on the intricate interplay between AI safety and AI security. Although the two disciplines now come together under the overarching goal of AI risk management, they have historically evolved separately, giving rise to differing perspectives. Therefore, in this paper, we advocate that stakeholders in AI risk management should be aware of the nuances, synergies, and interplay between safety and security, and unambiguously take into account the perspectives of both disciplines in order to devise mostly effective and holistic risk mitigation approaches. Unfortunately, this vision is often obfuscated, as the definitions of the basic concepts of "safety" and "security" themselves are often inconsistent and lack consensus across communities. With AI risk management being increasingly cross-disciplinary, this issue is particularly salient. In light of this conceptual challenge, we introduce a unified reference framework to clarify the differences and interplay between AI safety and AI security, aiming to facilitate a shared understanding and effective collaboration across communities.

Memorization in large language models (LLMs) is a growing concern. LLMs have been shown to easily reproduce parts of their training data, including copyrighted work. This is an important problem to solve, as it may violate existing copyright laws as well as the European AI Act. In this work, we propose a systematic analysis to quantify the extent of potential copyright infringements in LLMs using European law as an example. Unlike previous work, we evaluate instruction-finetuned models in a realistic end-user scenario. Our analysis builds on a proposed threshold of 160 characters, which we borrow from the German Copyright Service Provider Act and a fuzzy text matching algorithm to identify potentially copyright-infringing textual reproductions. The specificity of countermeasures against copyright infringement is analyzed by comparing model behavior on copyrighted and public domain data. We investigate what behaviors models show instead of producing protected text (such as refusal or hallucination) and provide a first legal assessment of these behaviors. We find that there are huge differences in copyright compliance, specificity, and appropriate refusal among popular LLMs. Alpaca, GPT 4, GPT 3.5, and Luminous perform best in our comparison, with OpenGPT-X, Alpaca, and Luminous producing a particularly low absolute number of potential copyright violations. Code can be found at https://github.com/felixbmuller/llms-memorization-copyright.

Deep learning models are widely used in traffic forecasting and have achieved state-of-the-art prediction accuracy. However, the black-box nature of those models makes the results difficult to interpret by users. This study aims to leverage an Explainable AI approach, counterfactual explanations, to enhance the explainability and usability of deep learning-based traffic forecasting models. Specifically, the goal is to elucidate relationships between various input contextual features and their corresponding predictions. We present a comprehensive framework that generates counterfactual explanations for traffic forecasting and provides usable insights through the proposed scenario-driven counterfactual explanations. The study first implements a deep learning model to predict traffic speed based on historical traffic data and contextual variables. Counterfactual explanations are then used to illuminate how alterations in these input variables affect predicted outcomes, thereby enhancing the transparency of the deep learning model. We investigated the impact of contextual features on traffic speed prediction under varying spatial and temporal conditions. The scenario-driven counterfactual explanations integrate two types of user-defined constraints, directional and weighting constraints, to tailor the search for counterfactual explanations to specific use cases. These tailored explanations benefit machine learning practitioners who aim to understand the model's learning mechanisms and domain experts who seek insights for real-world applications. The results showcase the effectiveness of counterfactual explanations in revealing traffic patterns learned by deep learning models, showing its potential for interpreting black-box deep learning models used for spatiotemporal predictions in general.

Artificial Intelligence (AI) models are now being utilized in all facets of our lives such as healthcare, education and employment. Since they are used in numerous sensitive environments and make decisions that can be life altering, potential biased outcomes are a pressing matter. Developers should ensure that such models don't manifest any unexpected discriminatory practices like partiality for certain genders, ethnicities or disabled people. With the ubiquitous dissemination of AI systems, researchers and practitioners are becoming more aware of unfair models and are bound to mitigate bias in them. Significant research has been conducted in addressing such issues to ensure models don't intentionally or unintentionally perpetuate bias. This survey offers a synopsis of the different ways researchers have promoted fairness in AI systems. We explore the different definitions of fairness existing in the current literature. We create a comprehensive taxonomy by categorizing different types of bias and investigate cases of biased AI in different application domains. A thorough study is conducted of the approaches and techniques employed by researchers to mitigate bias in AI models. Moreover, we also delve into the impact of biased models on user experience and the ethical considerations to contemplate when developing and deploying such models. We hope this survey helps researchers and practitioners understand the intricate details of fairness and bias in AI systems. By sharing this thorough survey, we aim to promote additional discourse in the domain of equitable and responsible AI.

What counts as legitimate AI ethics labor, and consequently, what are the epistemic terms on which AI ethics claims are rendered legitimate? Based on 75 interviews with technologists including researchers, developers, open source contributors, and activists, this paper explores the various epistemic bases from which AI ethics is discussed and practiced. In the context of outside attacks on AI ethics as an impediment to "progress," I show how some AI ethics practices have reached toward authority from automation and quantification, and achieved some legitimacy as a result, while those based on richly embodied and situated lived experience have not. This paper draws together the work of feminist Anthropology and Science and Technology Studies scholars Diana Forsythe and Lucy Suchman with the works of postcolonial feminist theorist Sara Ahmed and Black feminist theorist Kristie Dotson to examine the implications of dominant AI ethics practices. By entrenching the epistemic power of quantification, dominant AI ethics practices -- employing Model Cards and similar interventions -- risk legitimizing AI ethics as a project in equal and opposite measure to which they marginalize embodied lived experience as a legitimate part of the same project. In response, I propose humble technical practices: quantified or technical practices which specifically seek to make their epistemic limits clear in order to flatten hierarchies of epistemic power.

This paper aims to develop a semi-formal design space for Human-AI interactions, by building a set of interaction primitives which specify the communication between users and AI systems during their interaction. We show how these primitives can be combined into a set of interaction patterns which can provide an abstract specification for exchanging messages between humans and AI/ML models to carry out purposeful interactions. The motivation behind this is twofold: firstly, to provide a compact generalisation of existing practices, that highlights the similarities and differences between systems in terms of their interaction behaviours; and secondly, to support the creation of new systems, in particular by opening the space of possibilities for interactions with models. We present a short literature review on frameworks, guidelines and taxonomies related to the design and implementation of HAI interactions, including human-in-the-loop, explainable AI, as well as hybrid intelligence and collaborative learning approaches. From the literature review, we define a vocabulary for describing information exchanges in terms of providing and requesting particular model-specific data types. Based on this vocabulary, a message passing model for interactions between humans and models is presented, which we demonstrate can account for existing systems and approaches. Finally, we build this into design patterns as mid-level constructs that capture common interactional structures. We discuss how this approach can be used towards a design space for Human-AI interactions that creates new possibilities for designs as well as keeping track of implementation issues and concerns.

The use of artificial intelligence (AI) in research across all disciplines is becoming ubiquitous. However, this ubiquity is largely driven by hyperspecific AI models developed during scientific studies for accomplishing a well-defined, data-dense task. These AI models introduce apparent, human-recognizable biases because they are trained with finite, specific data sets and parameters. However, the efficacy of using large language models (LLMs) -- and LLM-powered generative AI tools, such as ChatGPT -- to assist the research process is currently indeterminate. These generative AI tools, trained on general and imperceptibly large datasets along with human feedback, present challenges in identifying and addressing biases. Furthermore, these models are susceptible to goal misgeneralization, hallucinations, and adversarial attacks such as red teaming prompts -- which can be unintentionally performed by human researchers, resulting in harmful outputs. These outputs are reinforced in research -- where an increasing number of individuals have begun to use generative AI to compose manuscripts. Efforts into AI interpretability lag behind development, and the implicit variations that occur when prompting and providing context to a chatbot introduce uncertainty and irreproducibility. We thereby find that incorporating generative AI in the process of writing research manuscripts introduces a new type of context-induced algorithmic bias and has unintended side effects that are largely detrimental to academia, knowledge production, and communicating research.

With the advances in computationally efficient artificial Intelligence (AI) techniques and their numerous applications in our everyday life, there is a pressing need to understand the computational details hidden in black box AI techniques such as most popular machine learning and deep learning techniques; through more detailed explanations. The origin of explainable AI (xAI) is coined from these challenges and recently gained more attention by the researchers by adding explainability comprehensively in traditional AI systems. This leads to develop an appropriate framework for successful applications of xAI in real life scenarios with respect to innovations, risk mitigation, ethical issues and logical values to the users. In this book chapter, an in-depth analysis of several xAI frameworks and methods including LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations) are provided. Random Forest Classifier as black box AI is used on a publicly available Diabetes symptoms dataset with LIME and SHAP for better interpretations. The results obtained are interesting in terms of transparency, valid and trustworthiness in diabetes disease prediction.

Researchers, practitioners, and policymakers with an interest in AI ethics need more integrative approaches for studying and intervening in AI systems across many contexts and scales of activity. This paper presents AI value chains as an integrative concept that satisfies that need. To more clearly theorize AI value chains and conceptually distinguish them from supply chains, we review theories of value chains and AI value chains from the strategic management, service science, economic geography, industry, government, and applied research literature. We then conduct an integrative review of a sample of 67 sources that cover the ethical concerns implicated in AI value chains. Building upon the findings of our integrative review, we recommend three future directions that researchers, practitioners, and policymakers can take to advance more ethical practices across AI value chains. We urge AI ethics researchers and practitioners to move toward value chain perspectives that situate actors in context, account for the many types of resources involved in co-creating AI systems, and integrate a wider range of ethical concerns across contexts and scales.