9607 research outputs found
Sort by
Towards A Modular End-To-End Machine Learning Benchmarking Framework
No full textMachine learning (ML) benchmarks are crucial for evaluating the performance, efficiency, and scalability of ML systems, especially as the adoption of complex ML pipelines, such as retrieval-augmented generation (RAG), continues to grow. These pipelines introduce intricate execution graphs that require more advanced benchmarking approaches. Additionally, collocating workloads can improve resource efficiency but may introduce contention challenges that must be carefully managed. Detailed insights into resource utilization are necessary for effective collocation and optimized edge deployments. However, existing benchmarking frameworks often fail to capture these critical aspects.We introduce a modular end-to-end ML benchmarking framework designed to address these gaps. Our framework emphasizes modularity and reusability by enabling reusable pipeline stages, facilitating flexible benchmarking across diverse ML workflows. It supports complex workloads and measures their end-to-end performance. The workloads can be collocated, with the framework providing insights into resource utilization and contention between the concurrent workloads
CoolWalks for active mobility in urban street networks
No full textWalking is the most sustainable form of urban mobility, but is compromised by uncomfortable or unhealthy sun exposure, which is an increasing problem due to global warming. Shade from buildings can provide cooling and protection for pedestrians, but the extent of this potential benefit is unknown. Here we explore the potential for shaded walking, using building footprints and street networks from both synthetic and real cities. We introduce a route choice model with a sun avoidance parameter α and define the CoolWalkability metric to measure opportunities for walking in shade. We derive analytically that on a regular grid with constant building heights, CoolWalkability is independent of α, and that the grid provides no CoolWalkability benefit for shade-seeking individuals compared to the shortest path. However, variations in street geometry and building heights create such benefits. We further uncover that the potential for shaded routing differs between grid-like and irregular street networks, forms local clusters, and is sensitive to the mapped network geometry. Our research identifies the limitations and potential of shade for cool, active travel, and is a first step towards a rigorous understanding of shade provision for sustainable mobility in cities
Human excreta recycling in Sweden: a PESTEL-SWOT framework analysis – Review
No full textSource-separating sanitation systems can maximise resource recovery from wastewater and mitigate the environmental impacts of conventional wastewater treatment plants, including eutrophication and climate change. This study conducts a comprehensive review of the literature on source-separating sanitation systems in Sweden, aiming to identify the challenges hindering their diffusion and potential expansion opportunities. Employing a rapid evidence synthesis approach, we extracted data from the Web of Science and supplemented findings through hand-searches in additional electronic databases. Of the 713 studies initially identified, 24 met our stringent inclusion criteria. The analysis was structured around a combined PESTEL (Political, Economic, Technical, Social, Environmental, Legal) and SWOT (Strengths, Weaknesses, Opportunities, Threats) framework to synthesise the existing body of work and discern main patterns. The findings underscore the untapped strengths in these technologies' potential in enhancing nutrient recovery and food security, in addition to reducing eutrophication and greenhouse gas emissions. The studies analysed reported Sweden's strengths in source separation, highlighting organisational diversity, market benefits, social acceptance, technological readiness, and nutrient recovery, all contributing to the SDGs and addressing challenges such as eutrophication and limited sanitation access. The primary challenges were identified as social and cultural taboos towards the recycling of human excreta, disbelief in its quality as a fertiliser, concerns about hazardous substances like pharmaceuticals, and a preference for using it to grow non-food crops. Our article main contribution lies in proposing 12 structured upscaling strategies addressing these barriers and leveraging the opportunities identified including policy measures to incentivise circular practices, building support through stakeholder engagement, updating building codes to require double piping, and enhancing municipal-utility cooperation. While grounded in Sweden, our study contributes to research on the broader shift towards sustainable food systems by leveraging internal strengths and external opportunities in circular wastewater systems
Hidden Layer Interaction: A Technique to Explore the Material of Generative AI
No full textThis pictorial describes the process of developing an interaction technique for directly engaging with the hidden layers of a generative AI model for image synthesis. First, we give some background to generative AI in HCI, arguing that current interaction techniques prevent us from directly interacting with the material of AI, foreclosing its use in design. Drawing on inspiration from the Computer Science field of feature visualization, we investigate the materiality of our prototype, a GAN model trained to generate fashion imagery, and show how Hidden Layer Interaction offers an alternative to standard prompting. In doing so, we illustrate how this change in approach leads to new forms of interaction with the internal semantics of generative AI, and demonstrate how one might use Hidden Layer Interaction to engage with AI as a material in design
Findings of the VarDial Evaluation Campaign 2025: The NorSID Shared Task on Norwegian Slot, Intent and Dialect Identification
No full textThe VarDial Evaluation Campaign 2025 was organized as part of the twelfth workshop on Natural Language Processing for Similar Languages, Varieties and Dialects (VarDial), colocated with COLING 2025. It consisted of one shared task with three subtasks: intent detection, slot filling and dialect identification for Norwegian dialects. This report presents the results of this shared task. Four participating teams have submitted systems with very high performance (> 97% accuracy) for intent detection, whereas slot detection and dialect identification showed to be much more challenging, with respectively span-F1 scores up to 89%, and weighted dialect F1 scores of 84%
Beyond Words: Personality-Aware Multimodal Modeling of Face-to-Face Emotion
Get PDFEmotions develop in layers: quick changes in valence and arousal form the core affect; when these are shaped by context and past experience, they become conscious feelings and are labeled as specific emotions. Language captures only parts of this process. More detailed cues are found in eye movements, pupil changes, facial action units (AUs), skin conductance, and EEG signals. Current emotion‑AI often overlooks these signals and ignores how personality shapes emotional responses, leading to models that are fragile and biased.This thesis approaches emotion recognition as a multimodal, trait‑aware task. Six studies range from controlled face‑viewing to simulated conversation. In 2 kHz eye‑tracking, just first fixations and microsaccades provide enough information for a simple multilayer perceptron to predict later gaze patterns in face emotion perception, showing that very fast eye movements are tied to emotion.Expanding on this, the public AFFEC dataset (5,073 trials, 73 participants) includes gaze, facial AUs, 64‑channel 256 Hz EEG, skin conductance, and Big‑Five personality scores, with each trial rated for both felt and perceived valence–arousal. Temporal mimicry analysis shows people copy fear expressions fastest, and extraverts are 20 ms quicker. Adding personality traits to gaze+face models cuts cross‑subject MSE by 18%The final model, MuMTAffect, a multitask transformer, reaches macro‑F1=0.55(valence) / 0.59 (arousal) and R2 > 0.94 for Big‑Five personality prediction, while keeping at least 85 % of its emotion accuracy even when one modality is missing.Key Contributions i. Introduction of the Advancing Face-to-Face EmotionCommunication (AFFEC) dataset, a comprehensive multimodal resource combining eye tracking, facial dynamics, Electroencephalography (EEG), and Galvanic Skin Response (GSR), enriched with both felt and perceived emotional labels and Big Five personality profiles. ii. Identification of microsaccades and facial mimicry timing as early, trait-sensitive markers of emotional response, offering high temporal precision and clear interpretability. iii. Integration of personality traits into model architecture, resulting in improved generalization and predictive accuracy, particularly for high-variance traits such as Neuroticism. iv. Development of the Multimodal Multitask AffectiveTransformer (MuMTAffect) Transformer, a multimodal, multitask model that fuses physiological signals and personality data to jointly predict emotions and stable traits with strong overall performance.Keywords: Affect, Emotion Recognition, Eye Tracking, Facial Action Units, Skin Conductance, EEG, Personality, Multimodal Fusion, Transformer Networks, valence–arousal
Leveraging and Probing Speech Prosody to Improve Spoken Language Processing
Get PDFNatural Language Processing (NLP) technologies have become ubiquitous, increasingly underpinning diverse applications across society. Although NLP advancements have predominantly focused on written text, especially with the recent emergence of powerful large language models (LLMs), spoken language remains fundamental to human communication. Speech inherently conveys richer and more nuanced information compared to text alone, encapsulating non-verbal and dynamic elements such as tone, pitch, intonation, pauses, and rhythm, collectively referred to as prosody. Prosody significantly influences the perception and intended meaning of spoken language, especially within tonal languages. Despite its importance, explicit utilization of prosodic features in spoken language processing remains limited, primarily due to the scarcity of high-quality prosodically annotated datasets and the lack of advanced models specifically designed to leverage such information. This thesis addresses these challenges through three significant contributions. First, it introduces the Akan Cinematic Emotions (AkaCE) dataset, the first-ever multimodal, prosodically annotated resource for an African language (Akan). AkaCE uniquely enables comprehensive modeling and analysis of prosody-informed Speech Emotion Recognition, establishing foundational benchmarks and facilitating future NLP research in under-represented African languages. Second, this thesis presents innovative computational models explicitly designed to integrate prosodic annotations, achieving state-of-the-art results across crucial Spoken Language Processing tasks, including Automatic Speech Recognition (demonstrating a Word Error Rate reduction of up to 28.3% on LibriSpeech) and speech-based instruction disambiguation in robotics (resolving ambiguous spoken commands with over 71% accuracy). Third, this work provides the first rigorous exploration of how multilingual contexts influence prosodic expression, revealing notable prosodic shifts in monolingual speech due to proximity to multilingual discourse. This finding highlights the potential of leveraging prosodic context as a powerful feature to enhance performance in complex NLP tasks such as code-switch detection. Empirical results presented throughout this thesis confirm that explicitly integrating speech prosody not only enhances task-specific model performance but also provides deeper linguistic insights across languages and discourse contexts, paving the way for more expressive, effective, and inclusive spoken language technologies
Summon a demon and bind it: A grounded theory of LLM red teaming
No full textEngaging in the deliberate generation of abnormal outputs from Large Language Models (LLMs) by attacking them is a novel human activity. This paper presents a thorough exposition of how and why people perform such attacks, defining LLM red-teaming based on extensive and diverse evidence. Using a formal qualitative methodology, we interviewed dozens of practitioners from a broad range of backgrounds, all contributors to this novel work of attempting to cause LLMs to fail. We focused on the research questions of defining LLM red teaming, uncovering the motivations and goals for performing the activity, and characterizing the strategies people use when attacking LLMs. Based on the data, LLM red teaming is defined as a limit-seeking, non-malicious, manual activity, which depends highly on a team-effort and an alchemist mindset. It is highly intrinsically motivated by curiosity, fun, and to some degrees by concerns for various harms of deploying LLMs. We identify a taxonomy of 12 strategies and 35 different techniques of attacking LLMs. These findings are presented as a comprehensive grounded theory of how and why people attack large language models: LLM red teaming
Towards A Modular End-To-End Machine Learning Benchmarking Framework
No full textMachine learning (ML) benchmarks are crucial for evaluating the performance, efficiency, and scalability of ML systems, especially as the adoption of complex ML pipelines, such as retrieval-augmented generation (RAG), continues to grow. These pipelines introduce intricate execution graphs that require more advanced benchmarking approaches. Additionally, collocating workloads can improve resource efficiency but may introduce contention challenges that must be carefully managed. Detailed insights into resource utilization are necessary for effective collocation and optimized edge deployments. However, existing benchmarking frameworks often fail to capture these critical aspects.We introduce a modular end-to-end ML benchmarking framework designed to address these gaps. Our framework emphasizes modularity and reusability by enabling reusable pipeline stages, facilitating flexible benchmarking across diverse ML workflows. It supports complex workloads and measures their end-to-end performance. The workloads can be collocated, with the framework providing insights into resource utilization and contention between the concurrent workloads
Beyond Isolated Factors: Investigating the Configurations That Shape User Responses to AI Advice
No full textAs artificial intelligence (AI) increasingly supports decision-making across various sectors, understanding how users perceive and respond to AI advice is critical. While previous studies have identified individual factors impacting this interaction between user and AI advice, these factors often interact in complex ways. This study explores how combinations of factors, including performance expectancy, effort expectancy, personal data use, and prediction explainability, influence user decisions to accept or reject AI advice. Using General Systems Theory as a framework and conducting experiments in a medical context with a Symptom Checker application, we aim to investigate these interdependencies through fuzzy-set qualitative comparative analysis. Our findings will contribute to the growing body of research on AI-advised decision-making by identifying key configurations that drive user acceptance or rejection of AI advice, offering insights both for academia and practical applications