Research Papers

ECQ$^{\text{x}}$: Explainability-Driven Quantization for Low-Bit and Sparse DNNs

Daniel Becking, Maximilian Dreyer, Wojciech Samek +2

The remarkable success of deep neural networks (DNNs) in various applications is accompanied by a significant increase in network parameters and arithmetic operations. Such increases in memory and computational demands make deep learning prohibitive for resource-constrained hardware platforms such as mobile devices. Recent efforts aim to reduce these overheads, while preserving model performance as much as possible, and include parameter reduction techniques, parameter quantization, and lossless compression techniques. In this chapter, we develop and describe a novel quantization paradigm for DNNs: Our method leverages concepts of explainable AI (XAI) and concepts of information theory: Instead of assigning weight values based on their distances to the quantization clusters, the assignment function additionally considers weight relevances obtained from Layer-wise Relevance Propagation (LRP) and the information content of the clusters (entropy optimization). The ultimate goal is to preserve the most relevant weights in quantization clusters of highest information content. Experimental results show that this novel Entropy-Constrained and XAI-adjusted Quantization (ECQ$^{\text{x}}$) method generates ultra low-precision (2-5 bit) and simultaneously sparse neural networks while maintaining or even improving model performance. Due to reduced parameter precision and high number of zero-elements, the rendered networks are highly compressible in terms of file size, up to $103\times$ compared to the full-precision unquantized DNN model. Our approach was evaluated on different types of models and datasets (including Google Speech Commands, CIFAR-10 and Pascal VOC) and compared with previous work.

Automated Machine Learning (AutoML) is a rapidly growing set of technologies that automate the model development pipeline by searching model space and generating candidate models. A critical, final step of AutoML is human selection of a final model from dozens of candidates. In current AutoML systems, selection is supported only by performance metrics. Prior work has shown that in practice, people evaluate ML models based on additional criteria, such as the way a model makes predictions. Comparison may happen at multiple levels, from types of errors, to feature importance, to how the model makes predictions of specific instances. We developed \tool{} to support interactive model comparison for AutoML by integrating multiple Explainable AI (XAI) and visualization techniques. We conducted a user study in which we both evaluated the system and used it as a technology probe to understand how users perform model comparison in an AutoML system. We discuss design implications for utilizing XAI techniques for model comparison and supporting the unique needs of data scientists in comparing AutoML models.

Artificial Intelligence (AI) systems exert a growing influence on our society. As they become more ubiquitous, their potential negative impacts also become evident through various real-world incidents. Following such early incidents, academic and public discussion on AI ethics has highlighted the need for implementing ethics in AI system development. However, little currently exists in the way of frameworks for understanding the practical implementation of AI ethics. In this paper, we discuss a research framework for implementing AI ethics in industrial settings. The framework presents a starting point for empirical studies into AI ethics but is still being developed further based on its practical utilization.

A new wave of decision-support systems are being built today using AI services that draw insights from data (like text and video) and incorporate them in human-in-the-loop assistance. However, just as we expect humans to be ethical, the same expectation needs to be met by automated systems that increasingly get delegated to act on their behalf. A very important aspect of an ethical behavior is to avoid (intended, perceived, or accidental) bias. Bias occurs when the data distribution is not representative enough of the natural phenomenon one wants to model and reason about. The possibly biased behavior of a service is hard to detect and handle if the AI service is merely being used and not developed from scratch, since the training data set is not available. In this situation, we envisage a 3rd party rating agency that is independent of the API producer or consumer and has its own set of biased and unbiased data, with customizable distributions. We propose a 2-step rating approach that generates bias ratings signifying whether the AI service is unbiased compensating, data-sensitive biased, or biased. The approach also works on composite services. We implement it in the context of text translation and report interesting results.

This late-breaking work presents a large-scale analysis of explainable AI (XAI) literature to evaluate claims of human explainability. We collaborated with a professional librarian to identify 18,254 papers containing keywords related to explainability and interpretability. Of these, we find that only 253 papers included terms suggesting human involvement in evaluating an XAI technique, and just 128 of those conducted some form of a human study. In other words, fewer than 1% of XAI papers (0.7%) provide empirical evidence of human explainability when compared to the broader body of XAI literature. Our findings underscore a critical gap between claims of human explainability and evidence-based validation, raising concerns about the rigor of XAI research. We call for increased emphasis on human evaluations in XAI studies and provide our literature search methodology to enable both reproducibility and further investigation into this widespread issue.

Youth are active users and stakeholders of artificial intelligence (AI), yet they are often not included in responsible AI (RAI) practices. Emerging efforts in RAI largely focus on adult populations, missing an opportunity to get unique perspectives of youth. This study explores the potential of youth (teens under the age of 18) to engage meaningfully in RAI, specifically through AI auditing. In a workshop study with 17 teens, we investigated how youth can actively identify problematic behaviors in youth-relevant ubiquitous AI (text-to-image generative AI, autocompletion in search bar, image search) and the impacts of supporting AI auditing with critical AI literacy scaffolding with guided discussion about AI ethics and an auditing tool. We found that youth can contribute quality insights, shaped by their expertise (e.g., hobbies and passions), lived experiences (e.g., social identities), and age-related knowledge (e.g., understanding of fast-moving trends). We discuss how empowering youth in AI auditing can result in more responsible AI, support their learning through doing, and lead to implications for including youth in various participatory RAI processes.

Deep learning models show promise in digital pathology, but their opaque decision-making processes limit trust and clinical adoption. To address this challenge, we present HIPPO, an explainable AI method for analyzing weakly-supervised multiple-instance learning (MIL) models that are widely used in whole slide image analysis. HIPPO constructs counterfactual whole slide images by systematically removing or adding selected tissue regions, providing a principled way to quantify how specific histologic areas influence model predictions under the MIL framework. This enables rigorous model interpretation, quantitative hypothesis testing, bias detection, and performance evaluation that extend beyond standard metrics. We demonstrate HIPPO across key clinical tasks, such as breast metastasis detection in axillary lymph nodes, prognostication in breast cancer and melanoma, and IDH mutation classification in gliomas. In metastasis detection, HIPPO uncovered critical model limitations undetectable by standard performance metrics or attention-based methods. For prognostic prediction, HIPPO outperformed attention by providing more nuanced insights into tissue elements influencing outcomes. In a proof-of-concept study, HIPPO facilitated hypothesis generation to identify melanoma patients who may benefit from immunotherapy. In IDH mutation classification, HIPPO more robustly identified the pathology regions responsible for false negatives compared to attention, suggesting its potential to outperform attention in explaining model decisions. In summary, HIPPO expands the explainable AI toolkit for computational pathology by enabling deeper insights into model behavior. This framework supports the trustworthy development, deployment, and regulation of weakly-supervised models in clinical and research settings, promoting their broader adoption in digital pathology.

How Are LLMs Mitigating Stereotyping Harms? Learning from Search Engine Studies

Alina Leidinger, Richard Rogers

With the widespread availability of LLMs since the release of ChatGPT and increased public scrutiny, commercial model development appears to have focused their efforts on 'safety' training concerning legal liabilities at the expense of social impact evaluation. This mimics a similar trend which we could observe for search engine autocompletion some years prior. We draw on scholarship from NLP and search engine auditing and present a novel evaluation task in the style of autocompletion prompts to assess stereotyping in LLMs. We assess LLMs by using four metrics, namely refusal rates, toxicity, sentiment and regard, with and without safety system prompts. Our findings indicate an improvement to stereotyping outputs with the system prompt, but overall a lack of attention by LLMs under study to certain harms classified as toxic, particularly for prompts about peoples/ethnicities and sexual orientation. Mentions of intersectional identities trigger a disproportionate amount of stereotyping. Finally, we discuss the implications of these findings about stereotyping harms in light of the coming intermingling of LLMs and search and the choice of stereotyping mitigation policy to adopt. We address model builders, academics, NLP practitioners and policy makers, calling for accountability and awareness concerning stereotyping harms, be it for training data curation, leader board design and usage, or social impact measurement.

Despite significant progress, evaluation of explainable artificial intelligence remains elusive and challenging. In this paper we propose a fine-grained validation framework that is not overly reliant on any one facet of these sociotechnical systems, and that recognises their inherent modular structure: technical building blocks, user-facing explanatory artefacts and social communication protocols. While we concur that user studies are invaluable in assessing the quality and effectiveness of explanation presentation and delivery strategies from the explainees' perspective in a particular deployment context, the underlying explanation generation mechanisms require a separate, predominantly algorithmic validation strategy that accounts for the technical and human-centred desiderata of their (numerical) outputs. Such a comprehensive sociotechnical utility-based evaluation framework could allow to systematically reason about the properties and downstream influence of different building blocks from which explainable artificial intelligence systems are composed -- accounting for a diverse range of their engineering and social aspects -- in view of the anticipated use case.

The rapid evolution of Large Language Models (LLMs) highlights the necessity for ethical considerations and data integrity in AI development, particularly emphasizing the role of FAIR (Findable, Accessible, Interoperable, Reusable) data principles. While these principles are crucial for ethical data stewardship, their specific application in the context of LLM training data remains an under-explored area. This research gap is the focus of our study, which begins with an examination of existing literature to underline the importance of FAIR principles in managing data for LLM training. Building upon this, we propose a novel framework designed to integrate FAIR principles into the LLM development lifecycle. A contribution of our work is the development of a comprehensive checklist intended to guide researchers and developers in applying FAIR data principles consistently across the model development process. The utility and effectiveness of our framework are validated through a case study on creating a FAIR-compliant dataset aimed at detecting and mitigating biases in LLMs. We present this framework to the community as a tool to foster the creation of technologically advanced, ethically grounded, and socially responsible AI models.

Machine Unlearning is rising as a new field, driven by the pressing necessity of ensuring privacy in modern artificial intelligence models. This technique primarily aims to eradicate any residual influence of a specific subset of data from the knowledge acquired by a neural model during its training. This work introduces a novel unlearning algorithm, denoted as Distance-based Unlearning via Centroid Kinematics (DUCK), which employs metric learning to guide the removal of samples matching the nearest incorrect centroid in the embedding space. Evaluation of the algorithm's performance is conducted across various benchmark datasets in two distinct scenarios, class removal, and homogeneous sampling removal, obtaining state-of-the-art performance. We also introduce a novel metric, called Adaptive Unlearning Score (AUS), encompassing not only the efficacy of the unlearning process in forgetting target data but also quantifying the performance loss relative to the original model. Additionally, we conducted a thorough investigation of the unlearning mechanism in DUCK, examining its impact on the organization of the feature space and employing explainable AI techniques for deeper insights.

Large Language Models (LLM) have made significant advances in the recent past becoming more mainstream in Artificial Intelligence (AI) enabled human-facing applications. However, LLMs often generate stereotypical output inherited from historical data, amplifying societal biases and raising ethical concerns. This work introduces i) the Multi-Grain Stereotype Dataset, which includes 52,751 instances of gender, race, profession and religion stereotypic text and ii) a novel stereotype classifier for English text. We design several experiments to rigorously test the proposed model trained on the novel dataset. Our experiments show that training the model in a multi-class setting can outperform the one-vs-all binary counterpart. Consistent feature importance signals from different eXplainable AI tools demonstrate that the new model exploits relevant text features. We utilise the newly created model to assess the stereotypic behaviour of the popular GPT family of models and observe the reduction of bias over time. In summary, our work establishes a robust and practical framework for auditing and evaluating the stereotypic bias in LLM.