서브메뉴
검색
상세정보
Relational Reasoning in Children and Machines: Insights Into Causal Generalization and Innovation.
Relational Reasoning in Children and Machines: Insights Into Causal Generalization and Innovation.
상세정보
- 자료유형
- 학위논문(국외)
- 기본표목-개인명
- 표제와 책임표시사항
- Relational Reasoning in Children and Machines: Insights Into Causal Generalization and Innovation.
- 발행, 배포, 간사 사항
- 발행, 배포, 간사 사항
- 형태사항
- 81 p.
- 일반주기
- Source: Dissertations Abstracts International, Volume: 87-04, Section: B.
- 일반주기
- Advisor: Gopnik, Alison.
- 학위논문주기
- Thesis (Ph.D.)--University of California, Berkeley, 2025.
- 요약 등 주기
- 요약Children excel at extracting deep relational structure from rather sparse data -identifying the underlying functional roles, causal mappings, and transformational rules that govern how objects and events relate-rather than relying on superficial co‐occurrence or perceptual similarity. They can further apply their learned structure to solve novel problems-a capacity that underlies tool use, analogy, and even scientific thinking. In this dissertation, I chart the developmental trajectory of three core relational-reasoning abilities and contrast them with what state-of-the-art AI systems achieve when trained on massive text and image corpora.In Chapter 1, preschoolers and adults move beyond mere associative pairings to solve tool‐substitution problems. In an "imitation" task, participants judge which object typically pairs with a tool (e.g. scissors with tape), tapping conventional co‐occurrences. In an "innovation" task, they must infer the causal affordance of the missing tool (e.g. selecting a bandage to stick torn paper), a leap that preschoolers achieve effortlessly while large language and vision models remain tethered to surface associations (Yiu et al., 2024).In Chapter 2, I introduce KiVA ("Kid‐inspired Visual Analogies"), a benchmark of visual analogy puzzles that probe five fundamental transformations-color, size, rotation, reflection, and number-grounded in classic developmental findings. Though three‐year‐old's readily infer and apply the abstract transformation rule (e.g. "rotate this object 90°"), state-of-the-art multimodal models (including "reasoning" models like GPT-o1) lag behind, especially when analogies demand counting or spatial inference beyond pixel‐level similarity (Yiu et al., 2025).In Chapter 3, I examine how humans actively build causal models. When presented with systems that vary in controllability and variability, children choose those that maximize their empowerment-the capacity to exert systematic, diverse control over outcomes. They are driven to discover structured relationships, not just any sort of variation or novelty (Yiu et al., in press).Together, these studies demonstrate that relational reasoning-the ability to abstract and manipulate functional and causal links-emerges early and underlies flexible tool use and innovation, analogy making, and causal generalization. In contrast, large multimodal models, which are trained primarily on large-scale text and image corpora with predictive and supervised learning algorithms, tend to capture only surface‐level co‐occurrences and perceptual similarities, lacking the rich causal structure and intrinsic exploration drives that characterize children's active, curiosity‐driven learning and genuine innovation.
- 주제명부출표목-일반주제명
- 주제명부출표목-일반주제명
- 주제명부출표목-일반주제명
- 비통제 색인어
- 비통제 색인어
- 비통제 색인어
- 비통제 색인어
- 비통제 색인어
- 비통제 색인어
- 부출표목-단체명
- 기본자료저록
- Dissertations Abstracts International. 87-04B.
- 전자적 위치 및 접속
- 원문정보보기
MARC
008260219s2025 us ||||||||||||||c||eng d■001000017359351
■00520260202105106
■006m o d
■007cr#unu||||||||
■020 ▼a9798297601550
■035 ▼a(MiAaPQ)AAI32236638
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a153
■1001 ▼aYiu, Eunice.
■24510▼aRelational Reasoning in Children and Machines: Insights Into Causal Generalization and Innovation.
■260 ▼a[S.l.]▼bUniversity of California, Berkeley. ▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a81 p.
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-04, Section: B.
■500 ▼aAdvisor: Gopnik, Alison.
■5021 ▼aThesis (Ph.D.)--University of California, Berkeley, 2025.
■520 ▼aChildren excel at extracting deep relational structure from rather sparse data -identifying the underlying functional roles, causal mappings, and transformational rules that govern how objects and events relate-rather than relying on superficial co‐occurrence or perceptual similarity. They can further apply their learned structure to solve novel problems-a capacity that underlies tool use, analogy, and even scientific thinking. In this dissertation, I chart the developmental trajectory of three core relational-reasoning abilities and contrast them with what state-of-the-art AI systems achieve when trained on massive text and image corpora.In Chapter 1, preschoolers and adults move beyond mere associative pairings to solve tool‐substitution problems. In an "imitation" task, participants judge which object typically pairs with a tool (e.g. scissors with tape), tapping conventional co‐occurrences. In an "innovation" task, they must infer the causal affordance of the missing tool (e.g. selecting a bandage to stick torn paper), a leap that preschoolers achieve effortlessly while large language and vision models remain tethered to surface associations (Yiu et al., 2024).In Chapter 2, I introduce KiVA ("Kid‐inspired Visual Analogies"), a benchmark of visual analogy puzzles that probe five fundamental transformations-color, size, rotation, reflection, and number-grounded in classic developmental findings. Though three‐year‐old's readily infer and apply the abstract transformation rule (e.g. "rotate this object 90°"), state-of-the-art multimodal models (including "reasoning" models like GPT-o1) lag behind, especially when analogies demand counting or spatial inference beyond pixel‐level similarity (Yiu et al., 2025).In Chapter 3, I examine how humans actively build causal models. When presented with systems that vary in controllability and variability, children choose those that maximize their empowerment-the capacity to exert systematic, diverse control over outcomes. They are driven to discover structured relationships, not just any sort of variation or novelty (Yiu et al., in press).Together, these studies demonstrate that relational reasoning-the ability to abstract and manipulate functional and causal links-emerges early and underlies flexible tool use and innovation, analogy making, and causal generalization. In contrast, large multimodal models, which are trained primarily on large-scale text and image corpora with predictive and supervised learning algorithms, tend to capture only surface‐level co‐occurrences and perceptual similarities, lacking the rich causal structure and intrinsic exploration drives that characterize children's active, curiosity‐driven learning and genuine innovation.
■590 ▼aSchool code: 0028.
■650 4▼aCognitive psychology.
■650 4▼aDevelopmental psychology.
■650 4▼aPsychology.
■653 ▼aAnalogical reasoning
■653 ▼aCausal learning
■653 ▼aChild cognition
■653 ▼aMultimodal AI models
■653 ▼aRelational reasoning
■653 ▼aTool use
■690 ▼a0633
■690 ▼a0620
■690 ▼a0621
■71020▼aUniversity of California, Berkeley▼bPsychology.
■7730 ▼tDissertations Abstracts International▼g87-04B.
■790 ▼a0028
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17359351▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
