DSpace@RPI (Rensselaer Polytechnic Institute)
Not a member yet
6809 research outputs found
Sort by
Knowledge graph enhanced large language models for food computing
No full textAugust2025School of ScienceRecent advances in large language models (LLMs) and the increasing availability of food-related data have led to significant progress in applying LLMs to food understanding. These developments have enabled Natural Language Processing (NLP) methods to address various food computing tasks, including food recognition, personalized recipe recommendation, and the generation of cooking guidelines. Despite the impressive performance and multi-modal adaptability of LLMs, domain-specific training remains essential for effective application. LLMs are still prone to hallucinations, outdated responses, and struggle in logical and numerical reasoning, limiting their utility in specialized domains. This thesis investigates supervised fine-tuning and instruction tuning of LLMs for text generation, and their use as a text processing engine for recommendation systems. Specifically, we explore their fine-tuning for recipe generation, nutritional estimation, and recipe data processing for recommendation tasks. We evaluate state-of-the-art small LLMs in the context of recipe generation and introduce LLaVA-Chef, a novel model trained using a multi-stage approach on a diverse dataset of recipe prompts. LLaVA-Chef significantly outperforms previous models, generating more detailed and accurate recipes with precise ingredient mentions — often surpassing the quality of human-authored recipes. Although prior research has highlighted hallucination issues in LLMs and explored incorporating contextual knowledge to improve factual accuracy, integration of food-specific knowledge graphs (KGs) with LLMs remains underexplored. To address this, we propose KERL, a unified system that leverages food KGs and LLMs to provide personalized food recommendations and generates recipes with associated micro-nutritional information. Given a natural language question, KERL extracts entities, retrieves subgraphs from the KG, which are then fed into the LLM as context to select the recipes that satisfy the constraints. Next, our system generates the cooking steps and nutritional information for each recipe. To evaluate our approach, we also develop a benchmark dataset by curating recipe related questions, combined with constraints and personal preferences. Through extensive experiments, we show that our proposed KG-augmented LLM significantly outperforms existing approaches, offering a complete and coherent solution for food recommendation, recipe generation, and nutritional analysis.Ph
Exploring formal methods for provably safe autonomous cyber-physical systems
No full textMay2025School of ScienceThe increasing complexity of autonomous cyber-physical systems (CPS) necessitates rigorous verification techniques to ensure safety and reliability. Safety-critical systems, such as those in aerospace, automotive, and medical domains, operate under strict constraints where failures can have catastrophic consequences. Modelling autonomous CPS as a hybrid system allows for verification of safety properties. Current verification methods, including model checking and Satisfiability Modulo Theories (SMT) solvers, face scalability challenges when dealing with large state spaces and complex non-linear dynamics. This thesis investigates theorem proving as a means to provide more scalable and explainable safety guarantees for cyber-physical systems. As a case study, we analyze the Mountain Car problem. We develop a discrete and continuous model for the dynamics. In this thesis we provide a formal proof using theorem proving and physical properties that shows that a simple controllerfor Mountain Car, under specific conjectures, ensures it reaches the goal position from any valid initial condition and without
a finite time bound for verification. The findings of this thesis highlight the potential for theorem proving to enhance the explainability and scalability of safety verification in autonomous cyber-physical systems.M
Testing platform and methodology for static characterization of gan hemt power semiconductor devices up to 650 v and 130 a across 50k to 400k
No full textMay2025School of EngineeringUnderstanding the behavior of power semiconductor devices at cryogenic temperatures is critical for NASA to enhance the reliability and performance of power electronics in space applications, where operation across extreme temperatures (50K to 400K) is required. Gallium Nitride (GaN) High Electron Mobility Transistors (HEMTs) are promising candidates for such environments. A static characterization platform has been developed to holistically evaluate and analyze V-I characteristics, reverse conduction, and leakage current for multiple samples concurrently across a broader temperature range. A key focus of this study is minimizing self-heating effects while optimizing pulse duration to balance accuracy and efficiency in characterizing GaN HEMTs. Understanding the behavior of power semiconductor devices at cryogenic temperatures is crucial for ensuring the reliability and efficiency of power electronics in space applications. Gallium Nitride (GaN) High Electron Mobility Transistors (HEMTs) have emerged as strong candidates for such extreme environments due to their high efficiency and fast switching capabilities. This study presents a static characterization platform capable of evaluating devices with voltage ratings up to 650 V and current ratings up to 130 A across a temperature range of 50K to 400K. The platform enables comprehensive analysis of I-V characteristics, reverse conduction, and leakage current, addressing challenges associated with self-heating effects through an enhanced static characterization method to optimize. A key aspect of this study is refining the characterization approach to ensure accuracy, repeatability, and efficient testing across multiple GaN HEMTs simultaneously. The developed static characterization platform consists of four core components: a cryocooler, matrix board, power device analyzer, and computer interface, each designed to minimize heat loss and electrical noise, maximizing measurement accuracy while minimizing the total automation testing time. The system architecture integrates automated test sequencing with an intuitive human-machine interface that processes data in real-time and generates plots for efficient analysis. This setup allows up to four power semiconductor devices to be tested concurrently, covering a wide range of electrical and thermal conditions. The methodology follows a structured testing approach to ensure the integrity of the device under test while preventing degradation and measurement inaccuracies. To validate the platform and methodology, case studies on the EPC 2307 and Infineon IGOT60R070D1 GaN devices were conducted, demonstrating successful characterization of forward conduction, reverse conduction, and leakage current. Results include detailed V-I characteristics across multiple gate voltages, on-state resistance trends, hysteresis analysis, and leakage current variations across the full temperature range. Furthermore, the experimental data for the IGOT60R070D1 device was directly compared to data from an independent research study, and the results showed a strong correlation, confirming the accuracy and reliability of the developed characterization platform. These findings provide a deeper understanding of how GaN devices behave under extreme thermal conditions, offering valuable insights for high-power applications. The developed platform represents a major advancement in cryogenic power semiconductor characterization, enabling faster, more efficient, and highly accurate testing of GaN devices rated up to 650 V and 130 A, setting a new benchmark for power electronics research and space applications.M
Multi-protocol mixing for multi-party computation
No full textAugust2025School of ScienceMulti-party computation (MPC) is a rapidly developing field of Computer Science and is approaching practical deployment; there are numerous backend implementations and protocol variants available for research use. This leaves users with a multitude of both backends and protocols to perform secure computation. Current backends give the user access to different MPC design paradigms as well as to different data representations. The choice of which MPC protocol is most efficient depends heavily on the computation being performed and requires the user to have extensive backend and protocol knowledge. Protocol Mixing is an optimization strategy that assigns protocols to computations in a way that maximizes the benefits of each protocol throughout an input program. The introduction of mixing algorithms would allow both experienced and new users to access the fastest possible execution times for their programs without needing to learn each backend and protocol they use. We present a toolchain which provides an automatic solution to the Optimal -Protocol Assignment problem, enabling developers to focus on designing their application logic without having to manage protocol selection and without requiring extensive knowledge of each backend. Our methodology comes with a provable guarantee; assuming structural properties which most practical programs meet, we show that the Optimal Protocol Assignment problem is tractable and provide a solution which demonstrates this claim. Our method leverages properties of SSA-form programs to divide the computation into manageable blocks, assign protocols to each statement within these blocks, discard provably sub-optimal blocks, and compute a merge schedule to yield a globally optimal set of assignments. Finally, we present results for a two-protocol and three-protocol backend (MP-SPDZ and MOTION respectively) for a standard set of benchmarks from protocol mixing literature. Notably, our solver produced assignments for Biometric Matching which were more efficient than the native mixer of MP-SPDZ, highlighting our mixer's ability to understand the global semantics of a program over locally optimal decisions.M
Investigating the dynamics of organic matter in the hudson river catchment across spatial, temporal, and hydrologic regimes
No full textAugust2025School of ScienceOrganic matter plays a critical role in the transfer of carbon between terrestrial, aquatic, and atmospheric reservoirs. As such, constraining the behavior and fluxes of organic matter in the environment is necessary to predict both short- and long-term exchange of carbon between these reservoirs. Environmental perturbations such as climate change and land use alterations have affected the production, mobilization, and fate of organic matter. However, many foundational studies have been performed on idealized, smaller scale systems leaving significant gaps in our understanding of organic matter sources and transport pathways. As the predictions of organic carbon fluxes between reservoirs are only as robust as our understanding of organic matter dynamics, there is a need to investigate previously unconstrained or understudied systems. This dissertation aims to improve our understanding of how historic and ongoing environmental perturbations influence the transport and storage of organic matter, and to identify previously overlooked source-inputs. In Chapter 1, I introduce the dynamics of organic matter, its role in global carbon budgets, and how its transport and fate can be altered by environmental perturbations. In Chapter 2, I assess the effects of hydrologic management, land use, and storm events on the quality of dissolved organic matter exported by the Hudson River. Given the region’s intensive hydrologic management and the anticipated increase in storm event frequency and intensity, this study provides insights into processes affecting similarly managed watersheds. I found that impoundments preferentially retain and degrade organic matter upstream, fundamentally changing the location and degree of organic matter processing. In Chapter 3, I assess whether the optical character of water extractable organic matter sequestered in the sediments of Brant Lake, NY effectively reflects historical environmental perturbations (e.g., logging, damming, and acid rain). Unlike rivers, which act as a conduit of organic matter transfer, lakes primarily serve as locations of carbon processing and storage. As such, understanding how environmental changes influence organic matter within lakes enhances our ability to interpret both past and future carbon dynamics in aquatic environments. In Chapter 4, I utilize compound-specific isotopic analysis to quantify shale-derived condensed aromatic compounds and estimate shale-derived bulk organic carbon inputs to the Hudson River. Although lithogenic sources have previously been excluded from short-term carbon cycling budgets, recent evidence suggests that rock-derived organic carbon may play more of an active role than previously assumed. By measuring this environmentally resistant fraction of organic matter, I estimated the contribution of petrogenic organic carbon and compared it to existing geochemical data. My findings indicate that shale-derived organic carbon is a plausible source of the aged organic carbon exported in the modern-day Mohawk and Hudson Rivers. While centered on the Hudson River system, the findings in this dissertation contribute to a growing understanding of how natural and anthropogenic forces shape the export and processing of organic carbon in aquatic systems.Ph
Thin film deposition of metastable face-centered cubic cobalt and ruthenium
No full textAugust2025School of EngineeringThe downscaling of devices leads to an increase in the resistivity of interconnect wires, causing resistance-capacitance delay. This has become a prominent roadblock in further advancement of microelectronics. The increase in the wire’s resistivity with decreasing line widths is primarily attributed to electron scattering from surfaces and gran boundaries. This effect is well explained by the Fuchs and Sondheimer and Mayadas and Shatzkez models for surface and grain boundary scattering, respectively. Theoretical predictions propose metastable face-centered cubic metals, such as cobalt and ruthenium, as potential replacements for copper, due to their predicted less pronounced resistivity increase at small dimensions. This motivates this study which delves into firstly synthesizing the metastable materials, followed by quantifying their electron transport. Firstly, I investigated synthesis of metastable fcc Co thin films. The phase composition of Co layers deposited by magnetron sputtering was studied as a function of processing gas (Ar or N2), temperature Ts = 100-600 °C, and substrate [Al2O3(0001), MgO(001) and SiO2/Si] in order to determine the kinetics for synthesis of metastable face-centered cubic (fcc) cobalt. N2 processing gas resulted in residual nitrogen in Co layers for Ts ≤ 200 °C but facilitated the growth of nitrogen-free fcc Co for Ts ≥ 300 °C. Templating with hexagonal Al2O3(0001) led to nucleation and growth of epitaxial hexagonal close-packed (hcp) Co(0001) in Ar. However, N2 caused stacking faults and the formation of a coherent mixed hcp/fcc epitaxial microstructure. Co on MgO(001) nucleated in the fcc phase, resulting in epitaxial fcc Co(001) layers in both Ar or N2 atmospheres. Density functional calculations confirmed the experimental observations, indicating that nitrogen facilitates formation of the cubic phase, with a predicted hcp-to-fcc transition at 10 at% N for T = 0 K and only 1.7 at.% N at 300 °C. They also suggest a negligible hcp(0001)/fcc(111) phase-boundary energy which facilitated the experimentally observed mixed hcp/fcc microstructure. The overall results demonstrates that the introduction of N2 gas during Co thin film deposition represents an effective approach to synthesize metastable fcc Co.
After recognizing the required kinetics barrier that led to formation of metastable fcc Co, I next quantified its electron transport within the vicinity of the Fuchs and Sondheimer model. In order to do this, I deposited face centered cubic Co thin films by reactive magnetron sputtering in 5 mTorr N2 at 400 °C followed by vacuum annealing at 500 °C. The resulting phase-pure Co(001)/MgO(001) layers contained negligible nitrogen, exhibited a surface roughness < 0.8 nm and a cube-on-cube epitaxial relationship with the substrate with Co[100] || MgO[100]. The measured resistivity vs thickness d = 10 - 1000 nm indicates a bulk resistivity ρo = 6.4 ± 0.3 µΩ-cm for fcc Co at room temperature and ρo = 1.3 ± 0.1 µΩ-cm at 77 K, and an effective electron phonon mean free path λ = 27 ± 2 nm and 79 ± 6 nm at 295 and 77 K, respectively. The resulting ρo × λ benchmark quantity is 3-5 times larger than predicted from first principles, suggesting a breakdown of the Fuchs-Sondheimer model at small dimensions. The overall results indicates that fcc Co exhibits no intrinsic conductance benefit over stable hcp Co nor conventional Cu for narrow interconnects.
Next, I investigated the effect of the phase change on magnetic properties of epitaxially grown cobalt layers of similar thicknesses. Fcc Co(001)/MgO(001) and hcp Co(0001)/Al2O3(0001) were epitaxially grown using an ultra-high vacuum magnetron sputtering system. X-ray diffraction analysis confirmed a single orientation epitaxial layer with negligible strain while a large mosaic spread particularly in the fcc layer. Atomic Force Microscopy (AFM) showed a rough topology for the hcp layer in comparison to the fcc, while Magnetic Force Microscopy (MFM) data shows clear large magnetic domains for the hcp layer (2 x 0.5 μm wide), while no domains are observed for fcc within the measured scan range. Surface Magneto-Optic Kerr Effect (SMOKE) measurements indicate coercivity for fcc and hcp Co as 48 Oe and 220 Oe respectively, suggesting a soft magnet like behavior for the fcc layer, in agreement with the MFM measurement. The measured saturation magnetization for fcc and hcp is 250 Oe and 500 Oe respectively. Magneto Resistance (MR) measurements on the two layers show a 0.4% change in Rs for the fcc layer while nearly 0.7% change for the hcp layers, suggesting electron scattering from magnetic domain boundaries is more pronounced for the hcp layers. This study portrays partial tuneability of the magnetic properties of Co via hcp to fcc transition.
Finally, I used the novel developed nitridation mechanism to synthesize another metastable fcc material, namely ruthenium. A combination of thin film deposition and first-principles calculations are employed to explore the effect of N2 gas on the phase formation of hexagonal close-packed (hcp) Ru, metastable face-centered cubic (fcc) Ru, and zincblende (zb) RuN. Sputter deposition in 20 mTorr Ar / N2 gas mixtures at 25 °C led to 1000-nm-thick phase-pure hcp Ru films for N2 partial pressures PN2 = 0 mTorr but a transition to fcc Ru and zb RuN with increasing PN2, indicating that nitrogen facilities the formation of metastable fcc Ru. This is confirmed by first-principles calculations on the Ru1-xNx formation energies which predict transitions from hcp Ru for x ≤ 0.077 to fcc Ru for 0.077 ≤ x ≤ 0.185 and zb Ru1-xNx for 0.204 ≤ x ≤ 0.500. This is in qualitative agreement the measured 0.073 ≤ x ≤ 0.241 for layers with phase-pure fcc Ru(N) and 0.241 ≤ x ≤ 0.496 for zb RuN. The fcc Ru lattice parameter afcc increases linearly with residual nitrogen content, with dafcc/dx = 0.009 Å/% and 0.01 Å/% from experiment and simulations, respectively. Deposition at higher temperature (Ts = 100 °C) led to no residual nitrogen within the Ru lattice and a phase change from zb (Ts = 25 °C) to hcp Ru (Ts = 100 °C). Alternatively, annealing experiments performed on the fcc Ru layer at Ta = 100 and 400 °C led complete loss of nitrogen by ta = 24 hours and 0.5 hour respectively, as well as a phase change from fcc to hcp Ru in both cases. The overall results demonstrate the use of nitrogen to facilitate synthesis of metastable fcc Ru.Ph
Characterizing liver pathology and innate immune activity in experimental models of simulated spaceflight, aging, and circadian disruption
No full textDecember 2024School of EngineeringHuman exploration in space is challenged by the exposure to environmental hazards that pose a risk to survival and recovery. The extent of physiological threats goes beyond what is currently known. Mechanosensitive bone resorption in microgravity is well documented and now focus has shifted to understanding how morphological changes in bone affect the bone marrow niche. Investigating bone marrow derived immune cells and their interaction in peripheral tissues is at the leading edge of spaceflight and disease research. This work uses models of environmental stresses relevant to spaceflight to deepen the understanding necessary to enable sustained human exploration in space. A motivating factor is the observed accumulation of lipids in the liver and apparent metabolic disorder of animals flown on the International Space Station. Others have highlighted the similarity to Non-Alcoholic Fatty Liver Disease (NAFLD), a prevalent condition on Earth, where innate immune response is a factor in recovery from or progression of the disease. Here we aim to recapitulate the observed condition by using mechanical unloading, radiation exposure, circadian disruption, and aging as individual factors experienced in spaceflight to deepen understanding of the threats posed. We find these models initiate early signs of NAFLD, though focus falls to altered monocyte function as a core feature of environmental disruption. Mechanical unloading and radiation distinctly influenced transcriptional signals of liver resident macrophages partially dependent on infiltrating monocytes. In our investigation of circadian disruption and neurodegenerative disease, monocytes in the bone marrow demonstrated distinct transcriptional regulation biasing them toward activation and inflammatory response. Finally in our model of aging we explore the function of cell fate regulator Cdkn1a/p21 and how it may shape monocyte response. Together these investigations take a multi-faceted exploratory approach to uncover how environmental stimuli may shape metabolic and immune health in space with applications to all humans on Earth.Ph
Optimization and generalization analysis of advanced neural networks and learning algorithms
No full textDecember 2024School of EngineeringIn recent years, deep learning has undergone rapid advancement. A notable trend in this development is the enhancement of learning efficiency for large foundation models, giving rise to numerous advanced learning algorithms designed for advanced neural networks. However, there remains a limited theoretical understanding of these algorithms and deep models. This thesis addresses this gap by delving into the optimization and generalization in advanced neural networks. The first part of the thesis focuses on the theoretical investigation of the basic block of advanced neural networks. The first work is to study one-layer single-head Vision Transformers (ViT), which is a self-attention layer followed by a two-layer perceptron. This work provides the sample complexity and the required number of iterations to achieve zero generalization on a binary classification task based on a data model where each data contains several label-relevant and label-irrelevant tokens. The sample complexity bound implies that a larger fraction of label-relevant tokens, a smaller token noise level, and a smaller initial model error can enhance the generalization. The theoretical finding also verifies the general intuition about the success of attention by proving that training using stochastic gradient descent (SGD) generates a sparse attention map focused on label-relevant tokens. Moreover, we also conclude that proper token sparsification can improve performance by removing label-irrelevant or noisy tokens, including spurious correlations. We then explore the generalization of Graph Transformer, a developing architecture originated from Transformers for graph learning. This work is based on a graph data model with discriminative nodes that determine node labels and non-discriminative nodes that are class-irrelevant. The theoretical results quantitatively characterize the sample complexity and number of iterations for convergence dependent on the fraction of discriminative nodes, the dominant patterns, and the fraction of erroneous labels. Meanwhile, we show that self-attention and positional encoding lead to generalization by making the attention map sparse and promoting the core neighborhood. This explains the superior feature representation of Graph Transformers compared with GCN and Graph Transformer without positional encoding. The second part of this thesis is to study modern algorithms on basic neural models. The first work studies learning with the group imbalance issue on a one-hidden-layer fully-connected neural network with a mixture of Gaussian input. This work quantifies the impact of individual groups on learning performance. The theoretical results include that when all group-level co-variance is in the medium regime and all mean are close to zero, we can achieve a small sample complexity, a fast training rate, and a high average and group-level testing accuracy. Moreover, it is shown that increasing the fraction of the minority group in the training data does not necessarily improve the generalization performance of the minority group. The second is the graph topology sampling with a three-layer Graph Convolutional Network (GCN), which not only consists of three-layer networks but also includes graph information in the model. This work characterizes sufficient conditions for graph topology sampling, such that GCN training leads to a diminishing generalization error on a semi-supervised node classification task. The sample complexity result explicitly depicts the impact of graph structures and topology sampling on the generalization performance. The third work is to study in-context learning (ICL), an inference method using pairs of testing data and labels as a prompt to make predictions without fine-tuning the model. We theoretically quantify the required number of training prompts and iterations and the length and distribution of the testing prompts for a desired ICL capability on unseen tasks with and without data distribution shifts. The training dynamics analysis also characterizes how different components in the learned Transformers contribute to the ICL performance. Moreover, this work proves that proper magnitude-based pruning has a minimal impact on performance while reducing inference costs. The last work is about Chain-of-Thought (CoT), a prompting method that incorporates multiple intermediate steps into each context example. This work establishes the first theoretical analysis of training nonlinear Transformers to obtain the CoT generalization ability by quantifying the required training samples and iterations. This work next theoretically characterizes the conditions for an accurate inference output by CoT when the provided reasoning examples contain noises and are not always accurate. Meanwhile, ICL, i.e., one-step CoT without intermediate steps, may fail to provide an accurate output when CoT does.Ph
Geometry-based and physics-informed 3d face & eye reconstruction for facial behavior analysis
No full textDecember 2024School of EngineeringFacial behavior analysis and recognition plays an important role in human-centered AI, boosting the technology development in the areas of human emotion recognition, attention detection and autonomous driving. This research performs 3D facial analysis, focusing on accurate 3D face and eye reconstruction, facial action recognition, and eye gaze estimation.The developments of deep learning models, combined with large benchmark datasets and representative 3D facial models, greatly improved the accuracy of 3D face and eye reconstruction. Despite this progress, existing methods in both areas still suffer from several significant limitations: a) lack of detailed shape modeling for accurately recovering subtle 3D facial motions and eyeball movement; b) over-dependence on a large amount of training data and labels; c) poor generalization across subjects and under different illuminations, distances and large head poses; and d) failure to effectively exploit physically plausible facial dynamics in videos. We introduce methods to address these limitations. For accurate 3D face reconstruction, we combine 3D facial models with Facial Action Unit (FAU) encoding system, where each AU represents a specific local facial motion driven by specific muscle activation.We first present a personalized 3D FAU blendshape learning framework together with a 3D face reconstruction model for recovering AU-interpretable 3D facial details. We also innovatively incorporate general knowledge of AU correlations into the learning process to reduce the amount of expression labels used in training. Our method not only produces a more personalized and detailed 3D face model but also yields improved facial action recognition. For 3D eye reconstruction, we create a deformable eye shape basis for representing detailed 3D eye structure. Different from existing approaches, we incorporate the 3D eye shape basis into a learning-based eye gaze estimation framework, inducing a geometry-based weak supervision in training the deep model. Our model is superior to others in terms of recovering 3D eye shape, eye rotation and gaze simultaneously from an image and is less dependent on full training labels while still maintaining the gaze accuracy. To further address the generalization and to exploit the facial dynamics for both facial actions and eye movement, we propose dynamic 3D face action and eye gaze tracking methods from monocular videos. The intuitive idea is based on the facial anatomy that all the facial motion components are activated by certain muscle contractions, so the reconstructed 3D motion should match with the physical laws of motion (Newton’s second law). Different from our frame-based method, we design different physically plausible models for facial action units and eyeball movement. For facial action units, we design a physics-informed model by constraining the reconstructed sequence to satisfy the underlying physics laws. For dynamic gaze tracking, we propose a physics-informed gaze tracking system by subjecting the eyeball movements to certain physical constrains and biomechanical laws. Furthermore, we propose to leverage human interactions and hand-eye coordination to reduce 3D eye gaze annotation using weakly supervised eye gaze tracking models. Our methods are evaluated against state-of-the-art methods both quantitatively and qualitatively, including 3D face reconstruction accuracy, facial action unit detection accuracy, and gaze estimation accuracy, both within and across datasets.Ph
Towards Computer-Using Personal Agents
No full textComputer-Using Agents (CUA) enable users to automate increasingly-complex tasks using graphical interfaces such as browsers. As many potential tasks require personal data, we propose Computer-Using Personal Agents (CUPAs) that have access to an external repository of the user's personal data. Compared with CUAs, CUPAs offer users better control of their personal data, the potential to automate more tasks involving personal data, better interoperability with external sources of data, and better capabilities to coordinate with other CUPAs in order to solve collaborative tasks involving the personal data of multiple users.
This report is a result of Dagstuhl Seminar 25051 "Trust and Accountability in Knowledge Graph-Based AI for Self Determination", which took place in January 2025Schloss Dagstuhl
Leibniz-Zentrum für Informatik Gmb