137091 research outputs found
Sort by
The Senators Problem: A Design Space of Node1 Placement Methods for Geospatial Network Visualization
No full textGeographic network visualizations often require assigning nodes to geographic coordinates, but this can be challenging when precise node locations are undefined. This thesis explores this problem using the U.S. senators as a case study. Each state has two senators making it difficult to assign clear individual locations. We devise eight different node placement strategies ranging from geometric approaches such as state centroids and longest axis midpoints to data-driven methods using population centers and home office locations. Through expert evaluation, we found that specific coordinates such as senators’ office locations and state centroids are preferred strategies, while random placements and the longest axis method are least favored. The findings also highlight the importance of aligning node placement with research goals and avoiding potentially misleading encodings. This thesis contributes to future advancements in geospatial network visualization software development and aims to facilitate more effective exploratory spatial data analysis.UndergraduateComputer Scienc
Value Expressions in Patents: Their Relationship to Patent Valuation and Technological Orientation
No full textThis dissertation examines the role and influence of value expressions (VEs)—defined as broad, optional, and narrative statements that highlight societal or commercial implications or potential benefits—in patent documents, focusing on their relationship to patent value and technological orientation. The study employs an explanatory sequential design that begins with a quantitative phase employing state-of-the-art natural language processing techniques. This includes using a generative language model, GPT-4o, to generate high-quality labeled examples, fine-tuning a discriminative language model, RoBERTa Large, on these examples, and deploying the fine-tuned model to classify sentences across the entire patent database. Regression analysis is then used to test hypotheses about the relationship between VEs, patent value, and technological orientation. The quantitative phase is followed by a qualitative analysis consisting of 12 semi-structured interviews comprising four inventors, five patent attorneys, two patent examiners, and one linguistics scholar, to explore the mechanisms guiding the observed relationships.
The study analyzes a database of 175,730 US artificial intelligence and nanotechnology patents and patent applications filed between 2005 and 2023, equivalent to approximately 7.1 million sentences. Patent IDs and metadata, including patent value indicators, were obtained from InnovationQ+ using expert-validated Boolean search strategies, while background and summary text sections, used to extract VEs, were downloaded from the USPTO's PatentsView database. The mapping of patents to UN Sustainable Development Goals (SDGs) based on their technology classification, used to create a measure of technological orientation with the categories socially oriented or not, was provided by LexisNexis Intellectual Property Solutions.
The research reveals significant findings about the prevalence and nature of VEs in patents. Analysis shows that 70% of patents contain private value expressions (PRIVEs, which highlight commercial implications), 59% have public value expressions (PVEs, which emphasize societal implications), and 24% include expressions combining both public and private values in a single statement (PVE-PRIVEs). The study finds a strong positive association between PVEs and traditional indicators of patent value, including forward citations (measuring technological impact), family size (indicating global reach), technology classes (reflecting technological breadth), and claims count (representing legal scope). This relationship holds across different regression models and remains robust in various sensitivity tests.
To assess whether VEs genuinely reflect the underlying technological potential of inventions or serve primarily as rhetorical devices, the research examines the relationship between these narrative elements and the technical features of patents as measured by their technology classification. Specifically, it analyzes the alignment between different types of VEs and the technological orientation of patents, operationalized through the mapping of technology classes to SDGs to identify socially oriented patents. The results demonstrate that patents containing PVEs show a positive association with social orientation, while PRIVEs show a negative association. Notably, patents containing sentences that include both public and private value expressions (PVE-PRIVEs) demonstrate the strongest positive relationship with social orientation.
Through interviews with patent attorneys, examiners, inventors, and a linguist, the study reveals that VEs in general, but more specifically PVEs, serve both normative and instrumental roles. While PVEs often originate from inventors' materials and provide a means for inventors to demonstrate their alignment with contemporary science, technology, and innovation norms, they are strategically placed by patent attorneys to enhance patent value through mechanisms such as demonstrating utility, avoiding admitting prior art, and broadening scope. The interviews helped conceptualize these dynamics through the notion of "societal value embedding," which explains how inventors and patent drafters incorporate narratives of societal impact into patents, simultaneously aligning with contemporary norms while potentially enhancing patent value.
The dissertation makes several key contributions. Methodologically, it introduces a novel approach using generative language models to analyze unstructured data and complex concepts in patent text at scale, and more generally abstract social science concepts in written text. Theoretically, it advances understanding of how and why value-based statements manifest in the patent system and relate to both patent value and technological orientation. The findings have significant implications for science, technology, and innovation policy, suggesting that patent documents can serve as valuable tools for anticipating and monitoring the societal impacts of emerging technologies. The study concludes by outlining policy implications and future research directions, emphasizing how patent offices could play a more active role in facilitating the analysis of societal impacts through structured approaches to documenting broader impacts in patent applications.Ph.D.Public Polic
Single-Cell Impedance Cytometry for the Control of Single-Cell Electroporation in a Flow-Through Device
No full textCell therapies have shown great promise in treating diseases such as CAR-T cells for blood cancers. Manufacturing these therapies presents considerable challenges and costs. These include quality of donor source material, transfection of cells, and controlling and measuring the quality of the product. Current therapies often transfect cells by viral vectors which are costly, have payload limitations, are difficult to target specific cells with, and present safety concerns due to immunogenicity and oncogenicity. A microfluidic single-cell electroporation device and scheme was made to address these concerns. Electroporation is an alternative non-viral method of transfection that is inexpensive and non-immunogenic. Additionally, it is applicable to different cell types or delivery payloads including larger payloads (e.g. CRISPR-Cas). However, electroporation can cause significant cell death, toxicity, and/or low delivery efficiencies. This microfluidic device electronically measures properties of single cells and applies an electroporation voltage based on that measurement. This feedback is demonstrated by selectively delivering to only certain cell types in a mixture of cell types. Further efforts were made to create a PDMS-on-glass version of the device to be compatible with microscopy imaging, as well as producing functional CAR-T cells with the device.Ph.D.Bioengineerin
Learning to See While Learning to Act: Diffusion Models for Active Perception in Robot Imitation
No full textRobots struggle with manipulating objects they can't fully observe, just as kids have trouble finding their toy hidden in a box. This paper introduces C3DM-Multiview, a novel approach that teaches robots to actively look around their environment while learning to complete tasks via diffusion models. Current methods for robot manipulation often rely on the observation from a fixed camera that can't capture occluded objects. C3DM-Multiview instead enables robots to dynamically adjust their viewpoints. Integrating camera movement with action refinement in a single diffusion model maintains a consistent relationship between what the robot sees and how the robot acts. Our camera-consistent noising strategy ensures that every noisy action in camera coordinates naturally emphasizes the pertinent dimensions for the current observations. C3DM-Multiview significantly outperforms existing methods on occlusion-specific tasks built based on Ravens benchmark, exceeding previous approaches by up to 72% on challenging scenarios where objects are initially occluded from view. When evaluated on the RLBench, our approach achieves a state-of-the-art 79% average success rate across six diverse manipulation tasks, surpassing methods that rely on full point-cloud based on observations from multiple fixed cameras. These results demonstrate that learning where to look while learning how to act leads to more robust robot manipulation capabilities, particularly in complex real-world environments.UndergraduateComputer Scienc
Hierarchical Stereo Navigation with Sparse Representation
No full textVision-based navigation systems with real-time requirements often employ hierarchical schemes that decompose navigation tasks across multiple temporal and spatial scales. This approach scales the problem down to real-time realizable sub-problems in local domains. Such hierarchical architectures typically consist of environment perception, planning, and control modules, which are arranged across three levels, from high to low. Leveraging sparse representations for these modules can reduce computational costs and offer favorable scalability across multiple devices. This is crucial for mobile robot navigation, where fast and safe decisions must be made when traversing unknown or partially known environments.
Different perception strategies create distinct levels of representation for vision-based navigation. Among the passive methods for constructing perception space, this thesis investigates depth estimation using lightweight stereo cameras. Specifically, it explores the use of sparse stereo representations in vision-based navigation. In the middle level, safety is a key concern in vision-based navigation. This thesis proposes gap-based local planning for various robot dynamics to ensure safety. Sparse gaps are detected within any laser scan-like egocentric perception. The low-level module in hierarchical navigation systems is motion control. Different from traditional pose-based control, this thesis describes a stereo image-based visual servoing system for trajectory tracking by nonholonomic robots without externally derived pose information nor a known visual map of the environment.
In summary, sparse representation is utilized across three levels—perception, motion planning, and motion control—in hierarchical stereo navigation. Sparse representation improves computational efficiency, safety, and task completion in stereo vision-based navigation systems.Ph.D.Mechanical Engineerin
Tools for interactive analysis of metagenomic read recruitment, sensitive detection and quantification of genes in metagenomes, and rapid, large scale genome relatedness estimation
No full textIn the modern era of genomics, nearly all genomic research ultimately flows through bioinformatic software. However, the rate at which biological sequence data is generated has outpaced increases in the availability of computational power. This has created an environment where most genome analysis programs must operate in the world of big data. Solving problems at large scales requires a fundamentally different approach and mindset than solving problems at smaller scales, and all modern bioinformatic software must be designed with solutions to this problem in mind. Software methods must be fast and efficient with computational resources to process modern datasets at all, and automation – which by necessity accounts for an ever-growing fraction of the process of genomic analysis – must be robust and able to translate increasingly complex and multifaceted data into clear, transparent, and comprehensible signal for researchers.
Multiple categories of solutions exist for these problems: software can be exceptionally quick and overcome scale with efficiency. Software can implement novel algorithms that achieve similar/identical goals to a previous approach, but which escape key limitations of the replaced approach. Software can limit the scope of hypothesis it seeks to render testable and eschew broad utility for use as a scalpel. Finally, software can summarize vast quantities of data into more digestible forms that are amenable to analyses which would not ordinarily scale well. In this thesis, we seek to address three software tools which explore this solution space and further the toolkit of genome analysis.
A common way to assess the relative abundance of a genome in a sample is by mapping reads from a corresponding metagenome to is (so-called, read recruitment). This analysis can also reveal is the reference is a good representative of the natural population in the sample, and if so, provides hints for population structure e.g., how clonal or diverse the population is. However, existing read recruitment plots are cumbersome to use, can visualize only one genome at a time, and are not interactive. In Chapter 2, we introduce RecruitPlotEasy, an interactive visualization program that allows for the rapid and detailed characterization of populations in metagenomes. RecruitPlotEasy addresses these previous limitations and empowers a researcher to manually curate many more populations than would otherwise be feasible and to communicate those results in a richly detailed manner.
A longstanding issue in metagenomics is the extraction of functional annotations from environments, using recovered genes to identify key functions. In some cases, this can even boil down to single key genes which are critical for the measurement of an important ecological process. In Chapter 3, we introduce ROCkI/O, a tool for the detection and quantification of target genes within metagenomes. ROCkI/O constructs gene-specific classification models which enable the rapid and extremely accurate retrieval of key genes – for example, environmentally important genes like denitrifiers or genes with biomedical significance such as antibiotic resistance genes – from metagenomic datasets. ROCkI/O takes the approach of adopting a narrow, but tailorable, scope in searching large collections of data, enabling searches which are both faster, require less resources, and are more sensitive than competing methods as long as the question being asked by researchers is sufficiently specific.
The comparison of novel genome sequences to those in public genome databases using genome similarity metrics such as whole genome aggregate amino acid identity (AAI) is a crucial step in determining the novelty and, if similar species exist, function of the novel genome. AAI functions to compare two genomes by averaging the amino acid similarity of their shared proteins, and has a long history of use as a method of funding distant relatives (genus or above). In Chapter 4, we introduce FastAAI v2, which improves upon the ultrafast AAI estimation capacities of the original version of FastAAI by increasing its robustness to large misestimations of AAI in prokaryotes and allows for the calculation of AAI in eukaryotes as well. In an era where the scale of public genome databases renders them difficult, slow, and expensive to search, FastAAI v2 overcomes the limitations of traditional AAI calculators through the use of a simple but effective, alignment-free metric of genome similarity. FastAAI v2 provides a scalable solution for the estimation of relatedness among genomes of intermediate novelty, including a proactive application to eukaryotic organisms which have been the targets of intense sequencing effort in recent years, and serves as a new, powerful tool in the toolkit of genome analysis.Ph.D.Bioinformatic
Tactile-augmented agent for slip-related dexterous manipulation
No full textContact-rich robotic grasping is crucial for manipulating objects in various industrial and household applications. In this thesis, we focus on a fluid collection task involving a multi-fingered robotic hand with 24 degrees of freedom and investigate the effects of incorporating tactile information into both sensing and the reward function. To this end, we designed and trained multiple reinforcement learning agents, each differing from the others in terms of how tactile information is incorporated into the sensing or reward signals. Specifically, we consider four agents: i) Baseline: no tactile information; ii) Tactile-Aware (TA): the ability to sense tactile information; iii) Tactile-Reward (TR): tactile information influencing rewards; iv) TS + TR: tactile information incorporating into both sensing and reward signals. Our results demonstrate that the ability to sense tactile information (i.e., the TS agent) can significantly improve the learning speed and final reward of the agents compared to the baseline agent. However, we found that adding tactile information to the reward function can decrease sample efficiency and lead to slower learning. Further, our qualitative analysis shows that the presence of tactile information and reward can affect the learned grasping gesture. Specifically, we find that leveraging tactile information for reward signals leads to more natural and ”human-like” grasps compared to agents that do not utilize shaped rewards that incorporate tactile information. Our study highlights the importance of considering tactile feedback in robotic grasping tasks and provides insights into the effects of different forms of tactile information and reward on learning performance.UndergraduateComputer Scienc
Reliable Decision-Making Under Uncertainty Through The Lens of Statistics and Optimization
No full textIn this thesis, we develop computationally efficient algorithms with statistical guarantees for problems of decision-making under uncertainty, particularly in the presence of large-scale, noisy, and high-dimensional data. In Chapter 2, we propose a kernelized projected Wasserstein distance for high-dimensional hypothesis testing, which finds the nonlinear mapping that maximizes the discrepancy between projected distributions. In Chapter 3, we provide an in-depth analysis of the computational and statistical guarantees of the kernelized projected Wasserstein distance. In Chapter 4, we study the variable selection problem in two-sample testing, aiming to select the most informative variables to determine whether two datasets follow the same distribution. In Chapter 5, we present a novel framework for distributionally robust stochastic optimization (DRO), which seeks an optimal decision that minimizes expected loss under the worst-case distribution within a specified set. This worst-case distribution is modeled using a variant of the Wasserstein distance based on entropic regularization. In Chapter 6, we incorporate Phi-divergence regularization into the infinity-type Wasserstein DRO, which is a formulation particularly useful for adversarial machine learning tasks. Chapter 7 concludes with an overview of promising future research directions.Ph.D.Industrial Engineerin
Smart Projectile Performance Augmentation Using an Articulating Nose
No full textThis dissertation examines the integration and performance of an articulating nose mechanism in projectile control systems. The articulating nose provides powerful minimum-phase control similar to a canard, generating substantial normal force for steering while adding no passive drag penalty when in its neutral position, unlike conventional control surfaces. The research employs a specialized multibody dynamic model alongside an articulating nose projectile linear theory to enable comprehensive analysis of articulating nose systems. The study demonstrates that an actively controlled articulating nose can effectively stabilize open-loop unstable projectiles, offering an alternative approach to traditional fin and spin stabilization methods. Maneuverability and agility characteristics of an articulating nose projectile are investigated alongside an examination of limitations such as roll coupling response. Further, the research explores how an articulating nose can extend projectile range through more favorable lift-to-drag ratios compared to fin-based control systems. Overall, this dissertation establishes the viability of the articulating nose as a control mechanism that can bring distinct benefits to stability, agility, maneuverability, and aerodynamic efficiency.Ph.D.Robotic
Optimizing Brain Stimulation for Parkinson's Disease, Memory Enhancement, and Optogenetic Control
No full textBrain stimulation has become an effective tool to treat drug-resistant patients with neurological diseases such as Parkinson’s disease, epilepsy, and many more emerging conditions. To achieve the greatest benefit, brain stimulation parameters – values such as the amplitude, frequency, and location of electrical current delivered to the brain – must be personalized according to each patient’s electrode placement, neuroanatomy, and physiological expression of their disease. Currently, clinicians must select each patient’s stimulation parameters manually by monitoring their symptoms (e.g. tremor severity, frequency of seizures) in a months-long trial-and-error process, delaying optimal clinical outcomes. Automating this process can lead to earlier improvement in patients’ quality-of-life and better outcomes for many applications of brain stimulation where manual parameter selection is challenging due to limited symptom expression. Here, I first describe a fully automated system for optimizing deep brain stimulation treatment for Parkinson’s disease, a well-established brain stimulation indication. This approach uses meta-active learning, an AI method that learns from a database of prior patients to (in real-time) interact with the brain and quickly find parameters that maximize neurophysiological biomarkers measuring which neural pathways are engaged by stimulation. Second, I examine the neurophysiological effects of cortical stimulation on brain-wide networks in patients with epilepsy, so that this methodology can be extended for patient-specific targeting of a novel brain stimulation paradigm for repairing memory dysfunction in diseases such as epilepsy and traumatic brain injury. Last, I examine how parameter optimization can enable precision modulation of hippocampal activity in rats using optogenetic stimulation (the control of neural activity using light) to improve the use of brain stimulation in translational research settings.Ph.D.Biomedical Engineerin