9607 research outputs found
Sort by
Modyn: Data-Centric Machine Learning Pipeline Orchestration
No full textIn real-world machine learning (ML) pipelines, datasets are continuously growing. Models must incorporate this new training data to improve generalization and adapt to potential distribution shifts. The cost of model retraining is proportional to how frequently the model is retrained and how much data it is trained on, which makes the naive approach of retraining from scratch each time impractical. We present Modyn, a data-centric end-to-end machine learning platform. Modyn's ML pipeline abstraction enables users to declaratively describe policies for continuously training a model on a growing dataset. Modyn pipelines allow users to apply data selection policies (to reduce the number of data points) and triggering policies (to reduce the number of trainings). Modyn executes and orchestrates these continuous ML training pipelines. The system is open-source and comes with an ecosystem of benchmark datasets, models, and tooling. We formally discuss how to measure the performance of ML pipelines by introducing the concept of composite models, enabling fair comparison of pipelines with different data selection and triggering policies. We empirically analyze how various data selection and triggering policies impact model accuracy, and also show that Modyn enables high throughput training with sample-level data selection
BECOMING IN/FERTILE: Data Practices and Reproductive Re-con-figurations
Get PDFThis dissertation explores how bodies, temporalities, and orientations become figured, configured, and reconfigured within everyday practices of fertility sensemaking. Fertility sensemaking refers to the ways people understand their bodies through data (e.g., about them, or in relation to statistics), alongside socio-cultural norms of reproduction and temporal scripts. Such practices become increasingly supported by various technologies that enter homes, clinics, and bodies, to generate and analyze data around reproductive bodies. It is thus necessary to understand how such data practices, and the technologies they enroll, augment and afford relations to, and understandings of, fertility. Throughout this dissertation I develop a qualitative analysis of fertility sensemaking that is grounded in interdisciplinary engagements with work in feminist theory, Human-computer interaction (HCI), and Science and Technology Studies (STS), and anchors in theories on posthumanism and crip/queer temporalities. I build on a range of empirical material, including bodily experiences around data obtained through mundane reproductive technologies, such as Menstruation and Fertility Tracking Applications (MFTAs), online forums, as well as medicalized datafication practices in Fertility Awareness Counseling (FAC), to scrutinize how different sites of datafication (the intimate, the shared, the medicalized) participate in the re-con-figuration of fertility. Rather than only being a ‘quality of the body’, this dissertation brings forth a conception of fertility as entangled, material, and relational practices. The three papers included in this dissertation contribute to HCI, STS, as well as feminist theory, and argue respectively 1) how reproductive bodies become figured through the datafication technologies; 2) how different objects and subjects come together, and configure fertile time and temporalities through relational and distributed practices of fertility sensemaking; and 3) how orientations to fertility become reconfigured in terms of possibility, time, and space, as infertility rather than fertility becomes anticipated
Does Random Movements mean Random Results? Why Asynchrony in Experiments on Body Ownership does not Work as Intended
No full textEffects of embodying virtual avatars are routinely validated experimentally by comparing synchronous and asynchronous movements between virtual and real bodies. This experimental paradigm, however, lacks justification, validation, and standardization. Asynchrony is implemented in numerous ways, such as through delayed, dislocated, or prerecorded movements, and these may impact embodiment and user experience distinctively. An online study () revealed that variations of asynchrony cause disparate responses to embodiment and user experience, with prerecorded movements distorting embodiment the most. A think-aloud study () revealed that asynchronous conditions lead to peculiar and oftentimes negative experiences. Furthermore, asynchronous conditions in some cases maintain, rather than break the body ownership illusion, as participants imitate the virtual body. Our results show that asynchrony in experiments on embodiment entails profound validity issues and should therefore be used with caution.Effects of embodying virtual avatars are routinely validated experimentally by comparing synchronous and asynchronous movements between virtual and real bodies. This experimental paradigm, however, lacks justification, validation, and standardization. Asynchrony is implemented in numerous ways, such as through delayed, dislocated, or prerecorded movements, and these may impact embodiment and user experience distinctively. An online study (N = 202) revealed that variations of asynchrony cause disparate responses to embodiment and user experience, with prerecorded movements distorting embodiment the most. A think-aloud study (N = 16) revealed that asynchronous conditions lead to peculiar and oftentimes negative experiences. Furthermore, asynchronous conditions in some cases maintain, rather than break the body ownership illusion, as participants imitate the virtual body. Our results show that asynchrony in experiments on embodiment entails profound validity issues and should therefore be used with caution
Differentially Private High-Dimensional Approximate Range Counting, Revisited.
No full textLocality Sensitive Filters are known for offering a quasi-linear space data structure with rigorous guarantees for the Approximate Near Neighbor search (ANN) problem. Building on Locality Sensitive Filters, we derive a simple data structure for the Approximate Near Neighbor Counting (ANNC) problem under differential privacy (DP). Moreover, we provide a simple analysis leveraging a connection with concomitant statistics and extreme value theory. Our approach produces a simple data structure with a tunable parameter that regulates a trade-off between space-time and utility. Through this trade-off, our data structure achieves the same performance as the recent findings of Andoni et al. (NeurIPS 2023) while offering better utility at the cost of higher space and query time. In addition, we provide a more efficient algorithm under pure ε-DP and elucidate the connection between ANN and differentially private ANNC. As a side result, the paper provides a more compact description and analysis of Locality Sensitive Filters for Fair Near Neighbor Search, improving a previous result in Aumüller et al. (TODS 2022)
Improving Reasoning Performance in Large Language Models via Representation Engineering
No full textRecent advancements in large language models (LLMs) have resulted in increasingly anthropomorphic language concerning the ability of LLMs to reason. Whether reasoning in LLMs should be understood to be inherently different is, however, widely debated. We propose utilizing a representation engineering approach wherein model activations are read from the residual stream of an LLM when processing a reasoning task. The activations are used to derive a control vector that is applied to the model as an inference-time intervention, modulating the representational space of the model, to improve performance on the specified task. We publish the code for deriving control vectors and analyzing model representations. The method allows us to improve performance on reasoning benchmarks and assess how control vectors influence the final logit distribution of a model via metrics such as KL divergence and entropy. We apply control vectors to Mistral-7B-Instruct and a range of Pythia models on an inductive, a deductive and mathematical reasoning task. We show that an LLM can, to a certain degree, be controlled to improve its perceived reasoning ability by modulating activations. The intervention is dependent upon the ability to reliably extract the model's typical state when correctly solving a task. Our results suggest that reasoning performance can be modulated in the same manner as other information-processing tasks performed by LLMs and demonstrate that we are capable of improving performance on specific tasks via a simple intervention on the residual stream with no additional training
MorSeD: Morphological Segmentation of Danish and its Effect on Language Modeling
No full textCurrent language models (LMs) mostly exploit subwords as input units based on statistical co-occurrences of characters. Adjacently, previous work has shown that modeling morphemes can aid performance for Natural Language Processing (NLP) models. However, morphemes are challenging to obtain as there is no annotated data in most languages. In this work, we release a wide-coverage Danish morphological segmentation evaluation set. We evaluate a range of unsupervised token segmenters and evaluate the downstream effect of using morphemes as input units for transformer-based LMs. Our results show that popular subword algorithms perform poorly on this task, scoring at most an F1 of 57.6 compared to 68.0 for an unsupervised morphological segmenter (Morfessor). Furthermore, evaluate a range of segmenters on the task of language modeling
Faster, Deterministic and Space Efficient Subtrajectory Clustering
No full textGiven a trajectory T and a distance Δ, we wish to find a set C of curves of complexity at most , such that we can cover T with subcurves that each are within Fréchet distance Δ to at least one curve in C. We call C an (,Δ)-clustering and aim to find an (,Δ)-clustering of minimum cardinality. This problem variant was introduced by Akitaya et al. (2021) and shown to be NP-complete. The main focus has therefore been on bicriteria approximation algorithms, allowing for the clustering to be an (, Θ(Δ))-clustering of roughly optimal size.We present algorithms that construct (,4Δ)-clusterings of (k log n) size, where k is the size of the optimal (, Δ)-clustering. We use (n³) space and (k n³ log⁴ n) time. Our algorithms significantly improve upon the clustering quality (improving the approximation factor in Δ) and size (whenever ∈ Ω(log n / log k)). We offer deterministic running times improving known expected bounds by a factor near-linear in . Additionally, we match the space usage of prior work, and improve it substantially, by a factor super-linear in n, when compared to deterministic results
How to Value Open Source Contributions? An Institutional Perspective from CERN
No full textSupplementary material for the paper "How to Value Open Source Contributions? An Institutional Perspective from CERN" submitted to the Software Engineering in Practice (SEIP) track of ICSE 2026. This contains the analysis code and the interview questions used that were used in this work
Sensing Place in HCI: Towards Immersive Digital Design Through Observational Notebook Techniques
No full textPlace-making in Human-Computer Interaction (HCI) remains a persistent research challenge, especially in contexts where technology intersects with culture, heritage, and socio-political dynamics. The process of creating, understanding, and engaging with "place" in digital and physical environments requires sensitivity to the local histories, sensory realities, and lived experiences of the people who inhabit those spaces. This challenge becomes particularly pronounced when designing for diverse communities across the African continent, where a place is often defined not only by physical landscapes but also by cultural narratives, ancestral ties, and communal practices. This workshop will address these challenges by exploring methods for designing interactive systems that support meaningful place-making in African contexts. A central component is the use of notebook techniques inspired by observational and narrative practices that capture the richness of place. Originating from disciplines such as film-making and as well as anthropology, these techniques emphasize attentiveness to sensory details, micro-interactions, and the emotional textures of everyday life promoting the details that are often missing from digital representations of the African places
Autonomous cartographies of Danish platform housecleaning
No full textDanish research on platform work has largely abstained from employing participatory methodologies and engaging migrant gig-workers to promote socially robust knowledge on the topic, while methodological nationalism premised on the figure of the citizen-worker has created many epistemological challenges regarding how knowledge on Nordic platform work has been produced to date (Andersen & Spanger 2024). This presentation introduces a methodological framework drawing on critical geography, for counter-mapping migrant platform housecleaners’ own spatio-temporal perceptions and aspirations vis-à-vis their participation in the Danish gig economy. This framework is meant to include and foreground migrants’ multiple situated approaches to gig work through participatory workshops conducting autonomous cartographies (cf. Counter Cartographies Collective et al. 2012). The goal of this approach is to defy singular definitions of platform labour in Denmark, while simultaneously producing knowledge from below and serving as a political intervention.Based on an ongoing project and a pilot counter-cartographic workshop facilitated by the authors in the spring of 2025, the presentation will discuss the strengths, limitations, ethical considerations and social/political potential of such a methodological approach, siding with literature stressing the need to highlight -primarily migrant- gig workers’ resistance practices (e.g., Bonini & Trere 2024). The introduced framework prioritizes migrant workers’ collective needs as foundational for meaningful counter-politics and proposes interactive counter-maps as a supportive tool to collective organizing. Moreover, the presentation contributes to the discussion on researchers’ critical reflections upon the extractivist aspects of participatory research.<br/