71446 research outputs found
Sort by
Mechanistic Control of Language Models
No full textThis dissertation talks about mechanistic control of large language models (LLM). The intuition is that, in order to rigorously control large language models, it is beneficial to first understand their inner working mechanisms. Even though a fully mechanistic understanding of language models is far from complete at the time this dissertation is written, I will demonstrate several cases where such understanding provides a useful surface for control. I argue that connecting interpretability to real-life applications--e.g., alignment and safety--offers advantages to the development of both fields.
I first position this methodological approach within a broader scientific context, discussing and systematizing the role of interpretability in artificial intelligence research. I then present four intensive case studies, each leveraging certain interpretability insights to control the behavior of language models.
The interpretability insights I explored span emergent world representations, representation of internal knowledge, phenomenon of attention decay, and representation of users. The demonstrated effects of control include targeted editing of the world model, reduced hallucination, improved stability of system prompts in long conversations, and better planning in multi-turn dialogue.Engineering and Applied Sciences - Computer Scienc
Discovering genomic stripes at high resolution
No full textThe 3.2 billion base pair human genome is 2 meters long laid end to end, yet fits into cells' nuclei around 5 microns in diameter on average. To accomplish this, the genome must fold into the nucleus without making knots. The folding structure of the genome is hierarchical and impacts key functionality of the cell. We currently have tools to assay genomic structure at coarse-grained resolution, but lack the tools to assay fine-grained structures, e.g. loops of DNA that form around genes.
Here I present a first-in-class method to assay fine-grained structures, particularly genomic stripes, across entire human genomes at high resolution, demonstrating its functionality on 1200bp resolution, 400bp resolution, and even 200bp resolution, a scale intractable to analyze with current state-of-the-art algorithms. In addition to superior functionality and detection of stripes, my tool also contributes at least 2x runtime and 10x memory load improvements to the current state-of-the-art algorithm. Additionally, I present a biophysics-based statistical method to annotate the significance of each stripe structure, which draws on the fractal globule model of DNA packing within the nucleus.
Preliminary analysis suggests that stripe regions are biologically relevant, potentially forming at key regions that interface with proteins in cells' nuclei to regulate genes. Gene regulation is a process fundamental for cell function, so these results are exciting, revealing the power of unbiased computational tools to reveal new biology.Computer Scienc
Metasurfaces as a Platform for Space-Division Multiplexing
No full textThe impending capacity crunch in optical fiber communications is driving the need for a new technological breakthrough in multiplexing technology. Multiplexing is the technique used to encode information on the five orthogonal degrees of freedom of the electromagnetic wave. These are the amplitude, phase, polarization, wavelength, and spatial mode of light. As current commercial communications have exhausted the first four of these degrees of freedom, space is set to be the next frontier for increasing the channel capacity of the optical fiber. Therefore, there is major interest in space-division multiplexing (SDM) and the new generation of optical devices that must facilitate its implementation.
The spatial multiplexer (MUX) is the optical device that bridges the gap between single-mode fibers and the fiber that will be the backbone of SDM. There are two possible types of fiber that can allow SDM: (1) multimode fiber and (2) multicore fiber. Both fibers have uses in long-haul submarine links to short-reach data centers. As a result, there has been intensive research on the design of spatial MUXs for both fiber types. However, the technologies used to build these MUXs fall short in efficiency, crosstalk, and scalability that hinder their adaptation as the platform that enables spatial MUXs.
In this dissertation, the metasurface is tested as a potential technology that could allow efficient and scalable spatial MUXs. An inverse design technique known as adjoint optimization is used to design the metasurfaces. The mathematical theory behind this method is developed for the multiplexing problem and is related to familiar phase-retrieval techniques which provides physical insight. Metasurface-based spatial MUXs are designed for both multimode fiber and multicore fiber. For the former, measured fidelity of converted modes are as high as 1.42 dB. For the latter, measured insertion losses and crosstalk are as low as 1.2 dB and -40 dB, respectively. The work presented here demonstrates the viability of metasurfaces as an ideal optical platform for SDM.Engineering and Applied Sciences - Applied Physic
Role of Periodontal Ligament Stem Cells in BLXA4-induced Periodontal Regeneration
No full textPeriodontal ligament stem cells (PDLSCs) maintain periodontal tissue homeostasis and play a critical role in tissue regeneration and the resolution of inflammation. They release Specialized Pro-resolving Lipid Mediators (SPMs) to regulate inflammation, but less is known about how SPMs stimulate PDLSCs to promote regeneration. This study aims to determine the role of Göttingen Miniature PDLSCs (mpPDLSCs) in Benzo-lipoxin A4 (BLXA4)- mediated periodontal regeneration.
The study was conducted on four Göttingen female minipigs (~30 kg; 12-36 months old). This study was designed as a single experimental arm for each interventional group, where each minipig had eight surgically created two-wall bony defects in all four posterior quadrants to establish the periodontal disease model. The interventional groups were allocated as Control, BLXA4, BLXA4+PDLSC, and PDLSCs. After surgical debridement, the groups receiving the BLXA4 treatment were administered 0.05 μg/defect in 18 μl of the BLXA4 lipid nanoparticle, while those receiving PDLSCs had 1x106 cells of the cell suspension following the surgical debridement and the control group received only surgical debridement. Bone regeneration was assessed by microCT. The mpPDLSCs were isolated from freshly extracted lateral incisors from the same host pig. Stem cell markers were characterized by using flow cytometry (CD11b, CD163, CD90, CD105). The multilineage differentiation of mpPDLSCs was assessed by osteogenic, adipogenic, and fibroblastic differentiation. Alizarin Red S staining was used to evaluate osteogenic differentiation; Oil-Red O staining was used for adipogenic differentiation; and Picro Sirius Red staining was used for fibroblastic differentiation after 21 days. A colony-forming unit-fibroblast (CFU-f) assay was utilized to assess cell proliferation using 0.5% crystal violet staining. Furthermore, the in vitro impact of BLXA4 on mpPDLSCs were investigated concerning cell viability along with SPM receptor expression (FPR2 and GPR32) using immunocytochemistry.
The isolated mpPDLSCs expressed mesenchymal stem cell markers and lacked hematopoietic stem cell markers. mpPDLSCs were multipotent and differentiated along osteogenic, adipogenic, and fibroblastic lineages with appropriate stimulation. Seeding 100 cells per well in a 6-well plate resulted in colony formation after 10 days. Additionally, mpPDLSCs exhibited enhanced proliferation when stimulated with BLXA4, with or without E. coli LPS stimulation. Moreover, BLXA4-stimulated mpPDLSCs expressed FPR2 and GPR32. The lipidomic results indicated that BLXA4 and PDLSCs exhibited distinct effects on the lipid mediator profiles, leading to differential modulation of pro-inflammatory and pro-resolving pathways. It is noteworthy that their combination did not invariably result in additive effects, suggesting intricate and potentially synergistic interactions between BLXA4 signaling and PDLSC-mediated mechanism. The microCT data, in conjunction with the qualitative histological results, demonstrated that all interventional groups exhibited a significantly improved periodontal regeneration (P.05) in comparison to debridement alone.
In summary, this study demonstrates the potential of mpPDLSCs in enhancing periodontal regeneration, alone or when stimulated with BLXA4. The findings indicate that BLXA4 promotes cell proliferation and stimulates the expression of receptors associated with inflammatory resolution (FPR2 and GPR32) in mpPDLSCs. The significant improvements in bone regeneration observed across all interventional groups underscore the therapeutic potential of combining SPMs with stem cell therapy. These results provide a strong foundation for a larger randomized clinical trial designed to further evaluate and translate SPM-enhanced stem cell therapy for periodontal regeneration in humans, potentially offering a novel approach to managing periodontal diseases.Oral Biolog
Comparative Effectiveness of Endovascular Therapies and Language-Based Equity in the Management of Non-Acute Subdural Hematomas
No full textNon‐acute subdural hematomas (NASDH) are intracranial hemorrhages that form between the dura mater and the arachnoid membrane, typically arising from disruption of bridging veins and expanding to displace brain tissue, causing significant neurological manifestations.(1) Their incidence increases with age‐related factors such as antithrombotic use and cerebral atrophy, and it is projected that by 2030, NASDH will become the most common cranial neurosurgical condition.(2)
More than two decades ago, Mandai et al. first described middle meningeal artery embolization (MMAE) as an intervention to disrupt the inflammatory mechanisms driving hematoma expansion and recurrence.(3) Since then, MMAE has attracted considerable attention in the fields of interventional radiology, neurology, and neurosurgery, as evidenced by the release of three simultaneous phase‐3 clinical trials supporting its use as an adjunctive therapeutic strategy. However, studies have historically pooled adjunctive and standalone MMAE cases together in a single study group. As a result, little comparative data on its effectiveness as a standalone treatment is available. Accordingly, our first project investigates whether patients treated exclusively with MMAE achieve fewer reintervention events, shorter hospitalizations, and reduced adverse events in comparison to those who undergo surgical evacuation alone.
Undeniably, biological factors and interventions are not the only determinants of patient outcomes.(4) Today, we have considerable evidence on how social and contextual features may impact outcomes of clinical relevance in many patient populations. This may be especially pertinent for NASDH,(5) which disproportionately impacts older adult males with multiple comorbidities, some of whom are immigrants with limited English familiarity. Building on previous research that linked language disparities to clinical outcomes in various neurosurgical populations,(6) our second project examines whether similar patterns emerged in NASDH patients treated at a single academic center over the past seven years, while also accounting for other clinical and sociodemographic influences.Medical Scienc
Public Transit as Linkage to Care
No full textWith the growing housing crisis and shelters at capacity, many of those who are
unhoused turn to public transit for shelter, warmth, safety, and access to essential
services. Many transit agencies in the United States employ policing in response to
homelessness. This can include citations, fines, confiscation of property, arrests, and
involuntary psychiatric commitment. These responses perpetuate the homelessnessjail cycle, displacement, traumatic experiences, and limited access to services. The
Massachusetts Bay Transportation Authority (MBTA) will serve as a case study given
its operation in the Greater Boston area, which has some of the highest rates of
homelessness in the country. This thesis identifies specific non-police interventions
and the MBTA locations where these strategies could be utilized to respond to and
ensure the safety of unhoused persons navigating the system.Department of Urban Planning and Desig
Hemp-Town
No full textHemp-Town, located in Ji Xi, China, is a speculative urban design project that positions hemp as the primary economic species, valued for its low environmental impact, high adaptability, and material versatility. This thesis reimagines sustainability within the design discipline through three methods: morphology, analogy/biology, and typology/topography.
Mapping oblique visualizes the economic paradigms of the hemp-based town using a color-coded system—orange, green, and purple—embedded in line drawings. Diagraming hemp translates hemp’s fractal growth pattern into a generative logic for urban transportation planning and redistribution design.Massing Ji Xi adopts Chinese characters into the architectural massing design as a formal generator.
The thesis critiques estate-based production city models and symbolic urbanism approaches. Rather, hemp directs the design process as both design tool and design object. Through this lens, Hemp-Town offers an integrated framework for rethinking landscape, architecture, and city-making.Department of Landscape Architectur
The role of novel long non-coding RNA TRIDENT in non-small cell lung cancer
No full textNon-small cell lung cancer (NSCLC) continues to be the leading cause of cancer mortality. While significant advancements have been made in its treatment and management, it remains incurable due to tumor recurrence. The molecular mechanisms driving NSCLC involving protein-coding genes have been extensively studied. However, systematically identifying and characterizing the regulatory factors involved in NSCLC can also provide invaluable insight into the pathways involved and aid in developing novel therapeutic and diagnostic strategies. Long non-coding RNAs (lncRNAs) are a class of non-coding regulatory RNAs that have been implicated in many human diseases. These RNAs, primarily through their secondary structure, employ diverse mechanisms to regulate most processes important for cell survival. Due to bioinformatic and technical limitations and a nascent understanding of functional RNA domains, most of these RNA molecules remain uncharacterized.
During my dissertation, I sought to use existing RNAseq databases to identify and characterize novel lncRNAs that show differential regulation in NSCLC. Through our screen, we identified several lncRNAs that have been previously described and some not. We focused on the novel lncRNAs and utilized online tools to filter out lncRNAs whose expression was significantly associated with patient prognosis. Next, we employed CRISPRi and antisense oligos (ASOs) to knock them down in cellular models of NSCLC to assess the effect on classical cancer phenotypes such as proliferation, invasion, and migration. Our analysis revealed several lncRNAs, yet uncharacterized, that significantly alter cancer phenotypes when their levels are perturbed. We delved deeper into the mechanism of one such lncRNA - RP11-462G12.1 or TRIDENT. We found that TRIDENT expression was significantly induced upon EGFR activation. Perturbing levels of TRIDENT highlighted its role in cell proliferation and drug resistance. Using ChIRP-MS analysis we identified TRIM28 as a protein interactor of TRIDENT. TRIDENT promotes phosphorylation of TRIM28 and knocking down TRIDENT leads to the accumulation of DNA damage in cancer cells via decrease TRIM28 phosphorylation. Altogether, our results reveal a novel molecular pathway in which TRIDENT regulates TRIM28 phosphorylation to promote tumor cell growth and drug resistance.
My work highlights the critical roles lncRNAs can play in modulating disease mechanisms and response to treatments. Their effect, though subtle, can have implications for treatment strategies for patients and aid in the development of better therapeutic targetsMedical Science
Reliable Grasping with Tactile Sensing
No full textFor robots to seamlessly integrate into everyday life and interact with various objects in unstructured environments like homes, they must exhibit extreme reliability. However, robots today often struggle to recover from errors, and understanding the root causes of these failures remains challenging. This thesis addresses robotic reliability in grasping and manipulation through the application of tactile sensing. Specifically, I have designed a highly instrumented robot hand, developed a stochastic friction model for grasp slip prediction, and employed tactile sensing to create physics-machine learning hybrid models. These models aim to predict the stability of anticipated grasping tasks and determine extrinsic contact locations. Our findings reveal that by modeling friction as a stochastic variable, we can quantify uncertainties in grasping and manipulation more accurately. For stability prediction, our hierarchical physics-machine learning hybrid approach proves effective in addressing performance, data size requirements, interpretability, and generalizability. Overall, reliability is crucial for robots to effectively assist humans in various settings, including hospitals, elderly care facilities, and disaster sites. This thesis offers a comprehensive framework for understanding the role of tactile sensing in enhancing reliability prediction in robotic grasping and manipulation.Engineering and Applied Sciences - Engineering Science
Visual analytics at the atlas scale for multimodal and spatial single-cell data
No full textScientific measurements at the resolution of individual cells – single-cell experiments – are central to biology because the cell is the fundamental unit of life. In the past decade, advancements in sequencing and bioimaging technologies have enabled cellular measurements to be made at high-throughput, forging the field of single-cell biology. Single-cell experiments are now being applied in large scale, driven by concerted efforts from funding institutions and consortia worldwide. The resulting datasets are being compiled into single-cell atlas resources: collections of cellular- resolution maps intended to summarize and communicate single-cell data through hierarchical and spatial organization and inclusion of multiple biosamples, tissue types, organs, and/or organisms from one or more experimental conditions. Single-cell atlases are intended to facilitate downstream usage in biology and medicine via their establishment as gold-standard sets of measurements that can serve as common points of reference. Current information visualization systems are not equipped to adequately deal with the scale, complexity, and heterogeneity of single-cell atlas data, nor are they tailored to the needs of target user audiences.
This thesis investigates how visual analytics systems can be employed for interactive visualization of multimodal and spatial single-cell datasets, addressing challenges at the scale of individual experiments to whole atlases. The first chapter provides an overview of the landscape of single-cell data visualizations, including interactive systems. The second chapter builds on preliminary work on a framework for interactive data visualization for single-cell data, including for spatial, imaging, and multimodal data. The third chapter builds upon this framework to develop a system tailored to cross-experiment comparisons, for example between single-cell data from case and control groups, informed through interviews with individuals from its intended audience of biologists, pathologists, and clinicians. The fourth chapter explores how algorithms for identification of spatial domains with relevance to biological function can be integrated with interactive visualizations. The fifth chapter explores how to integrate chromatin accessibility measurements into the existing framework for single-cell data visualization.
By combining and extending concepts from bioinformatics, information visualization, human-computer interaction, and software engineering, this thesis pioneers approaches for exploring, understanding, and communicating foundational single-cell atlas resources.Medical Science