TY - BOOK

T1 - Beyond predictions: alignment between prior knowledge and machine learning for human-centric augmented intelligence

AU - Chen, Xia

PY - 2024/9/12

Y1 - 2024/9/12

N2 - As artificial intelligence (AI) and machine learning (ML) continue to achieve new advancements across a variety of domains, it is clear that data-driven, end-to-end, connectionist approaches can match and sometimes surpass human performance in many tasks. However, while these systems have achieved significant performance milestones, they fundamentally differ from human intelligence, especially in logic, symbolism, and reductionist mindsets. In design, engineering, and scientific research, this often results in two methodological developments running in parallel: data-driven methods and first-principles approaches, each with its unique strengths, natural limitations, and performance gaps. Inherently, both methodologies are capable of proficiently describing phenomena, or what we call modeling, differently yet without any intrinsic mutual exclusivity, mirroring the long-standing debate between connectionism and symbolism in AI. I recognize that separate modeling in either methodology is not a sustainable approach for long-term development. Instead, the key is to reconcile and integrate our prior knowledge into ML methods for engineering human-centric alignment. This integration is crucial in empirical-dominant domains, niche fields, and scientific explorations that align with logical positivism.We are facing an era where machine intelligence is reshaping our understanding of the world and our position within it. Recognizing human and machine's unique strengths and limitations is essential in harnessing both combined potentials. This dissertation proposes a machine assistance framework for human intelligence augmentation, primarily aiding users such as engineers, designers, and researchers during the decision-making process as an informative extension. The core effort is to establish the alignment between human insights and machine capabilities from cognitive, methodological, and interactional perspectives to surpass the limitations inherent in each separated methodology while utilizing information more efficiently and comprehensively. In this context, this dissertation is structured around three distinct yet essential objectives: Decision-Making Processes Alignment, Methodological Paradigms Alignment, and Communication Alignment. Each tackles real-world specific problems as case studies to demonstrate the practical application of the proposed paradigm and methodologies. The first alignment focuses on developing fundamental mechanisms to synchronize machine learning models with complex user decision-making processes. Drawing inspiration from the human nervous system and cognitive estimation processes, this framework integrates key mechanisms for identifying and quantifying uncertainty, while incorporating incomplete input in a dynamic, human-in-the-loop circumstance. It lays a groundwork for aligning information flow from different knowledge-based and data-driven methods to interact with users while enabling a recursive multi-objective optimization, showcasing its practical informational assistance utility in the decision-making process under uncertainty. The second alignment emphasizes methodological integration by proposing a paradigm to systematically embed human prior knowledge and domain insights into data-driven methods. The paradigm first identifies naturally inherited uncertainties from data acquisition conditions, data-driven model mechanisms, and prior domain knowledge. These identifications set the foundation for underscoring their complementary roles in information representation. Building upon this, I organize three hierarchical knowledge integration levels named "Ladder of Knowledge-integrated Machine Learning," corresponding to the three stages in advancing data-driven models with respect to data augmentation, modeling process enhancement, and knowledge discovery. These knowledge types are: direct modeling knowledge for system description (level 1), inductive logic and disentanglement (level 2), and abstract reasoning and deductive logic (level 3). With applications in engineering case studies analysis, I affirm the framework's efficacy in interpolation, extrapolation, and information representation tasks. With the ladder ascendance, the methodological paradigm in ML shifts from pattern finding to model building to describe a system, revealing key factors in ML that align with the human cognitive and learning process. The last alignment of this dissertation extends the topic to the human-computer interaction (HCI) domain to investigate an ecology of symbiosis between humans and machine assistance, exchanging information among phenomena, data, and prior knowledge, and how such interaction can be enhanced through diverse patterns and processing methods. Furthermore, I primitively investigate potential possibilities of how data-driven methods decode implicit information, such as electroencephalograms, to broaden the bandwidth of information exchange between humans and machines. To conclude, this dissertation aims to lay the groundwork for aligning human intelligence and machine capacities to advance human-centric intelligence augmentation, while providing a more holistic understanding of the interaction between human intelligence and computational methods.

AB - As artificial intelligence (AI) and machine learning (ML) continue to achieve new advancements across a variety of domains, it is clear that data-driven, end-to-end, connectionist approaches can match and sometimes surpass human performance in many tasks. However, while these systems have achieved significant performance milestones, they fundamentally differ from human intelligence, especially in logic, symbolism, and reductionist mindsets. In design, engineering, and scientific research, this often results in two methodological developments running in parallel: data-driven methods and first-principles approaches, each with its unique strengths, natural limitations, and performance gaps. Inherently, both methodologies are capable of proficiently describing phenomena, or what we call modeling, differently yet without any intrinsic mutual exclusivity, mirroring the long-standing debate between connectionism and symbolism in AI. I recognize that separate modeling in either methodology is not a sustainable approach for long-term development. Instead, the key is to reconcile and integrate our prior knowledge into ML methods for engineering human-centric alignment. This integration is crucial in empirical-dominant domains, niche fields, and scientific explorations that align with logical positivism.We are facing an era where machine intelligence is reshaping our understanding of the world and our position within it. Recognizing human and machine's unique strengths and limitations is essential in harnessing both combined potentials. This dissertation proposes a machine assistance framework for human intelligence augmentation, primarily aiding users such as engineers, designers, and researchers during the decision-making process as an informative extension. The core effort is to establish the alignment between human insights and machine capabilities from cognitive, methodological, and interactional perspectives to surpass the limitations inherent in each separated methodology while utilizing information more efficiently and comprehensively. In this context, this dissertation is structured around three distinct yet essential objectives: Decision-Making Processes Alignment, Methodological Paradigms Alignment, and Communication Alignment. Each tackles real-world specific problems as case studies to demonstrate the practical application of the proposed paradigm and methodologies. The first alignment focuses on developing fundamental mechanisms to synchronize machine learning models with complex user decision-making processes. Drawing inspiration from the human nervous system and cognitive estimation processes, this framework integrates key mechanisms for identifying and quantifying uncertainty, while incorporating incomplete input in a dynamic, human-in-the-loop circumstance. It lays a groundwork for aligning information flow from different knowledge-based and data-driven methods to interact with users while enabling a recursive multi-objective optimization, showcasing its practical informational assistance utility in the decision-making process under uncertainty. The second alignment emphasizes methodological integration by proposing a paradigm to systematically embed human prior knowledge and domain insights into data-driven methods. The paradigm first identifies naturally inherited uncertainties from data acquisition conditions, data-driven model mechanisms, and prior domain knowledge. These identifications set the foundation for underscoring their complementary roles in information representation. Building upon this, I organize three hierarchical knowledge integration levels named "Ladder of Knowledge-integrated Machine Learning," corresponding to the three stages in advancing data-driven models with respect to data augmentation, modeling process enhancement, and knowledge discovery. These knowledge types are: direct modeling knowledge for system description (level 1), inductive logic and disentanglement (level 2), and abstract reasoning and deductive logic (level 3). With applications in engineering case studies analysis, I affirm the framework's efficacy in interpolation, extrapolation, and information representation tasks. With the ladder ascendance, the methodological paradigm in ML shifts from pattern finding to model building to describe a system, revealing key factors in ML that align with the human cognitive and learning process. The last alignment of this dissertation extends the topic to the human-computer interaction (HCI) domain to investigate an ecology of symbiosis between humans and machine assistance, exchanging information among phenomena, data, and prior knowledge, and how such interaction can be enhanced through diverse patterns and processing methods. Furthermore, I primitively investigate potential possibilities of how data-driven methods decode implicit information, such as electroencephalograms, to broaden the bandwidth of information exchange between humans and machines. To conclude, this dissertation aims to lay the groundwork for aligning human intelligence and machine capacities to advance human-centric intelligence augmentation, while providing a more holistic understanding of the interaction between human intelligence and computational methods.

U2 - 10.15488/17976

DO - 10.15488/17976

M3 - Doctoral thesis

CY - Hannover

ER -