Pixels and Weights : The Situated Work of Teaching Computers to See
Engdahl, Isak LU
(2025)
In Lund Dissertations in Sociology
- Abstract
- The uses of artificial intelligence are increasing across many areas of contemporary society, including computer vision systems involving machine learning and artificial neural networks. These systems rearrange action possibilities and are often described as “black boxes” that challenge conventional modes of scientific understanding and engineering control. It is important to understand the activities through which such systems are constructed. The aim of this dissertation is to analyze the situated work involved in constructing computer vision components and systems: how research and development take shape in practice. Drawing on ethnographic fieldwork and interviews in computer vision research labs, the dissertation analyzes how computer... (More)
- The uses of artificial intelligence are increasing across many areas of contemporary society, including computer vision systems involving machine learning and artificial neural networks. These systems rearrange action possibilities and are often described as “black boxes” that challenge conventional modes of scientific understanding and engineering control. It is important to understand the activities through which such systems are constructed. The aim of this dissertation is to analyze the situated work involved in constructing computer vision components and systems: how research and development take shape in practice. Drawing on ethnographic fieldwork and interviews in computer vision research labs, the dissertation analyzes how computer vision scientists make computers capable of ‘seeing’. Adopting a pragmatist–interactionist perspective, the dissertation treats science and engineering as hybrid forms of collective action where scientists interact with computational objects and research apparatuses. The analysis follows three lines of their work: model development, the construction of evaluative benchmark datasets, and the implementation of models in new contexts. The first study argues that daily laboratory activity is organized through articulation work and pipeline welding, where computational components are integrated into operational models. These models become epistemically opaque yet meaningful through shared interpretive work in relationally sustained awareness contexts. The second study analyzes the creation of a benchmark dataset, showing how standards and protocols depend on less-visible alignment work to hold cooperation together. The third study investigates trajectories in which scientists adapt pre-trained models to local contexts, and develops the idea of a social license to capture how researchers coordinate expertise and effort when reconstructing external models for local use. This social license helps actors manage trajectories and navigate the contingencies that arise as models are adapted to new settings. The dissertation advances an interactionist and processual understanding of computational knowledge production, foregrounding the collective work that makes computer vision systems function beyond narratives of exceptionalism, crisis, or autonomy. (Less)
- Abstract (Swedish)
- Användningen av artificiell intelligens ökar inom många områden i det samtida samhället, inklusive datorsynssystem som bygger på maskininlärning och artificiella neuronnätverk. Dessa system omformar handlingsmöjligheter och beskrivs ofta som ”svarta lådor” som utmanar etablerade former av vetenskaplig förståelse och teknisk kontroll. Det är därför viktigt att förstå de aktiviteter genom vilka sådana system konstrueras. Syftet med denna avhandling är att analysera det situerade arbete som är inblandat i att konstruera komponenter och system för datorsyn – hur forskning och utveckling tar form i praktiken. Med utgångspunkt i etnografiskt fältarbete och intervjuer i forskningslabb för datorsyn undersöker avhandlingen hur forskare inom området... (More)
- Användningen av artificiell intelligens ökar inom många områden i det samtida samhället, inklusive datorsynssystem som bygger på maskininlärning och artificiella neuronnätverk. Dessa system omformar handlingsmöjligheter och beskrivs ofta som ”svarta lådor” som utmanar etablerade former av vetenskaplig förståelse och teknisk kontroll. Det är därför viktigt att förstå de aktiviteter genom vilka sådana system konstrueras. Syftet med denna avhandling är att analysera det situerade arbete som är inblandat i att konstruera komponenter och system för datorsyn – hur forskning och utveckling tar form i praktiken. Med utgångspunkt i etnografiskt fältarbete och intervjuer i forskningslabb för datorsyn undersöker avhandlingen hur forskare inom området gör datorer kapabla att ”se”. Utifrån ett pragmatistiskt–interaktionistiskt perspektiv behandlar avhandlingen vetenskap och ingenjörskonst som hybrida former av kollektivt handlande, där forskare interagerar med datorbaserade objekt och forskningsapparaturer. Analysen följer tre linjer i deras arbete: modellutveckling, konstruktionen av utvärderande referensdataset (benchmark-dataset) och implementeringen av modeller i nya kontexter. Den första studien visar hur det dagliga laboratoriearbetet organiseras genom artikulationsarbete och så kallad pipeline welding, där beräkningskomponenter integreras till fungerande modeller. Dessa modeller blir epistemiskt opaka men samtidigt meningsfulla genom gemensamt tolkningsarbete i relationellt upprätthållna awareness contexts. Den andra studien analyserar skapandet av ett benchmark-dataset och visar hur standarder och protokoll är beroende av mindre synligt alignment work för att hålla samarbetet samman. Den tredje studien undersöker de trajektorier där forskare anpassar förtränade modeller till lokala sammanhang, och utvecklar idén om en social license för att fånga hur forskare koordinerar expertis och insatser när externa modeller rekonstrueras för lokal användning. Denna sociala licens hjälper aktörerna att hantera trajektorier och navigera de kontingenser som uppstår när modeller anpassas till nya miljöer. Avhandlingen utvecklar ett interaktionistiskt och processuellt synsätt på produktionen av beräkningsbaserad kunskap och synliggör det kollektiva arbete som får datorsynssystem att fungera – bortom berättelser om exceptionalism, kris eller autonomi. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/record/2f559e73-96cb-482a-b968-f50e2743209d
- author
- Engdahl, Isak
LU
- supervisor
- opponent
-
- Associate Professor Ziewitz, Malte, Cornell University
- organization
- publishing date
- 2025-11
- type
- Thesis
- publication status
- published
- subject
- keywords
- lab ethnography, situated work, Artificial Intelligence (AI), Computer Vision, Machine Learning (ML), model development, benchmark datasets, model implementation, Science and Technology Studies (STS)
- in
- Lund Dissertations in Sociology
- issue
- 143
- pages
- 176 pages
- publisher
- Lund University
- defense location
- LUX Aula, Nedre. Helgonavägen 3, 223 62 Lund
- defense date
- 2025-11-28 13:15:00
- ISSN
- 1102–4712
- ISBN
- 978-91-8104-691-5
- 978-91-8104-692-2
- project
- Show & Tell: Scientific representation, algorithmically generated visualizations, and evidence across epistemic cultures
- Machine Pedagogy
- language
- English
- LU publication?
- yes
- id
- 2f559e73-96cb-482a-b968-f50e2743209d
- date added to LUP
- 2025-10-31 09:43:52
- date last changed
- 2025-11-03 13:25:11
@phdthesis{2f559e73-96cb-482a-b968-f50e2743209d,
abstract = {{The uses of artificial intelligence are increasing across many areas of contemporary society, including computer vision systems involving machine learning and artificial neural networks. These systems rearrange action possibilities and are often described as “black boxes” that challenge conventional modes of scientific understanding and engineering control. It is important to understand the activities through which such systems are constructed. The aim of this dissertation is to analyze the situated work involved in constructing computer vision components and systems: how research and development take shape in practice. Drawing on ethnographic fieldwork and interviews in computer vision research labs, the dissertation analyzes how computer vision scientists make computers capable of ‘seeing’. Adopting a pragmatist–interactionist perspective, the dissertation treats science and engineering as hybrid forms of collective action where scientists interact with computational objects and research apparatuses. The analysis follows three lines of their work: model development, the construction of evaluative benchmark datasets, and the implementation of models in new contexts. The first study argues that daily laboratory activity is organized through articulation work and pipeline welding, where computational components are integrated into operational models. These models become epistemically opaque yet meaningful through shared interpretive work in relationally sustained awareness contexts. The second study analyzes the creation of a benchmark dataset, showing how standards and protocols depend on less-visible alignment work to hold cooperation together. The third study investigates trajectories in which scientists adapt pre-trained models to local contexts, and develops the idea of a social license to capture how researchers coordinate expertise and effort when reconstructing external models for local use. This social license helps actors manage trajectories and navigate the contingencies that arise as models are adapted to new settings. The dissertation advances an interactionist and processual understanding of computational knowledge production, foregrounding the collective work that makes computer vision systems function beyond narratives of exceptionalism, crisis, or autonomy.}},
author = {{Engdahl, Isak}},
isbn = {{978-91-8104-691-5}},
issn = {{1102–4712}},
keywords = {{lab ethnography; situated work; Artificial Intelligence (AI); Computer Vision; Machine Learning (ML); model development; benchmark datasets; model implementation; Science and Technology Studies (STS)}},
language = {{eng}},
number = {{143}},
publisher = {{Lund University}},
school = {{Lund University}},
series = {{Lund Dissertations in Sociology}},
title = {{Pixels and Weights : The Situated Work of Teaching Computers to See}},
url = {{https://lup.lub.lu.se/search/files/231817401/Isak_Engdahl_-_Pixels_and_Weights.pdf}},
year = {{2025}},
}