DSpace@RPI (Rensselaer Polytechnic Institute)
Not a member yet
6809 research outputs found
Sort by
Deadlock freedom in actor languages
No full textDecember 2023We introduce a framework using session types for denoting the relationships between multiple actors in a system. We prove that ascribing to this framework guarantees deadlock freedom, and demonstrate tests of how an implementation of such would work in the SALSA language.M
A noise audit of human-labeled benchmarks for machine commonsense reasoning
No full textWith the advent of large language models, evaluating and benchmarking these systems on important AI problems has taken on newfound importance. Such benchmarking typically involves comparing the predictions of a system against human labels (or a single ‘ground-truth’). However, much recent work in psychology has suggested that most tasks involving significant human judgment can have non-trivial degrees of noise. In his book, Kahneman suggests that noise may be a much more significant component of inaccuracy compared to bias, which has been studied more extensively in the AI community. This article proposes a detailed noise audit of human-labeled benchmarks in machine commonsense reasoning, an important current area of AI research. We conduct noise audits under two important experimental conditions: one in a smaller-scale but higher-quality labeling setting, and another in a larger-scale, more realistic online crowdsourced setting. Using Kahneman’s framework of noise, our results consistently show non-trivial amounts of level, pattern, and system noise, even in the higher-quality setting, with comparable results in the crowdsourced setting. We find that noise can significantly influence the performance estimates that we obtain of commonsense reasoning systems, even if the ‘system’ is a human; in some cases, by almost 10 percent. Labeling noise also affects performance estimates of systems like ChatGPT by more than 4 percent. Our results suggest that the default practice in the AI community of assuming and using a ‘single’ ground-truth, even on problems requiring seemingly straightforward human judgment, may warrant empirical and methodological re-visiting
The effects of dynamic lighting and nature soundscapes on human well-being, perceived restorativeness, and cognitive function
No full textAugust 2024School of ArchitectureThis research extends on previous studies on acoustic and lighting on forest bathing to measure human well-being, cognitive effects and the perceived restorativeness of two experimental stimuli, soundscapes and lighting, presented to individuals indoors. Recent studies have shown that forest bathing has been proven to have multiple positive benefits on human health and well-being, which in itself has brought novel interest to explore how to reintroduce the forest bathing in urbanized indoor environments. Currently, there is no solid experimental foundation to suggest that recreation of nature through synthetic parameters indoors can positively affect individuals in well-being, cognitive function, and perceived restorativeness. This study is the implementation of lighting and acoustic interventions in an augmented indoor environment to stimulate lighting and sound conditions as if encountered in a typical wooded area to determine if these changes, likewise to forest bathing, could improve well-being, cognitive function, and perceived restorativeness with heart rate and heart rate variability testing, cognitive tasks, self-report stress and mood scales, and a perceived restorativeness scale. Results found that there was statistical significance only in the effect of time for heart rate, in the interaction between stimulus and group for State-Trait Inventory for Cognitive and Somatic Anxiety (STICSA) and the Digit Symbol Substition Task (DSST), and in the main effect of stimulus and group and stimulus interaction for the D2 test. Survey results showed a leaning preference towards the Nature condition to having restorative value while the Traffic and Nothing conditions did notM
Deep learning based orthognathic surgical planning
No full textNovember 2023Orthognathic surgery, which addresses issues with the jaw and face, is an intricate procedure that requires meticulous planning. Computer-Assisted Surgical Simulation (CASS) provides surgeons with a robust platform to refine their surgical strategies through simulated practice before the actual intervention, culminating in a detailed plan that guides the surgical correction. Despite the advancements, accurately crafting this plan is a significant challenge due to the reliability of reference model estimation and the speed of facial biomechanical simulations, as well as the iterative nature of plan revisions. These simulations aid surgeons in fine-tuning their approach, ensuring optimal functional and aesthetic outcomes through a cycle of evaluation and modification. Nonetheless, CASS has shown limitations in the efficiency and precision of the surgical plans which directly affect surgical outcomes. This thesis proposes the integration of deep learning techniques to overcome these limitations. Delving into the relatively untapped potential of deep learning for orthognathic surgery planning, this research confronted three major challenges: 1) How to develop a patient-specific reference bony model leveraging a normal subject dictionary for reliable surgical planning? 2) How can deep learning model the complex non-linear relationship between face and bone to expedite the facial simulation process? 3) How to directly target the facial soft-tissue during the planning phase to remove the need for repeated adjustments and to streamline outcome-focused planning? Overcoming these barriers is essential for a deep learning method that is not only technologically advanced but also clear, efficient, reliable, and clinically relevant. The structure of this dissertation is anchored around three specific aims: Our first aim focuses on the development of a self-supervised learning framework anchored in a deep query network. This framework utilizes a dictionary of normal subjects to facilitate the development of a dependable surgical plan customized for each patient. This innovative approach significantly diminishes initial planning time and heightens plan quality, streamlining the surgeon's workflow. The second aim enhances the precision and speed of facial simulation by integrating the power of deep learning with traditional surgical biomechanical simulation techniques like Finite Element Methods (FEM). This enhancement is realized through the introduction of a cutting-edge attentive correspondence assisted movement transformation network (ACMT-Net) that captures and models the complex non-linear relationship between the bone framework and facial tissues. The final aim shifts the planning paradigm. Instead of primarily looking at bone structures to devise the surgical plan, greater emphasis is placed on the anticipated postoperative appearance of the patient's face. This shift ensures that the surgical plan is formulated just once, automatically validated, and consistent with the patient’s aesthetic goals. This thesis outlines a pathway that adeptly circumvents the fundamental limitations of conventional CASS. It emphasizes the importance of achieving a patient's optimal facial appearance after surgery by incorporating the robust capabilities of deep learning into CASS. This resultant methodology produces a surgical planning methodology that is transparent, accurate, and adaptable, meeting modern surgical needs. This innovative approach aims to revolutionize orthognathic surgical planning by merging cutting-edge deep learning technology with a concentrated emphasis on post-operative aesthetic outcomes.Ph
Spatial location of binaural signals using cepstral analysis
No full textDecember 2023As described by the precedence effect, the time delay between a sound (lead) and reflection (lag) influences our ability to perceive the lead's spatial location. When the lead and lag overlap we will perceive them as a single auditory event called a binaural fusion. Binaural fusions merge the spatial characteristics of the lead and lag making it difficult to identify their origins. When the sound signal is periodic, like music, the potential for binaural fusions to occur increases dramatically. Precedence effect models of auditory perception have traditionally avoided binaural fusion by using noise signals or impulsive signals, by pre-calculating the signal's impulse response, or by calculating the impulse response after the fact from a discrete signal. Unfortunately such models are not useful in real-world scenarios as the impulse response is rarely known beforehand, periodic and impulsive signals often coexist, and sound is continuous. With the increased interest in spatial audio technologies comes an increased demand for precedence effect models that can be applied to real-world applications. In this thesis we present the cepstral binaural model (CEPBIMO), a perceptual model of the precedence effect that is more resilient to periodicity, reflections, and binaural fusions than existing models. In addition to exhibiting improved performance, CEPBIMO surpasses the abilities of existing models in that it can be applied to variable signal types including periodic signals, it does not require prior knowledge of the impulse response, and it can be applied to a running audio signal in real-time. The innovations of CEPBIMO stem from its novel application of cepstral analysis, a signal processing technique used to identify frequencies of periodicity in a signal. From the results of cepstral analysis a deconvolution filter is created and used to separate sounds from reflections. For a given binaural signal CEPBIMO will return a binaural activity map, a visual representation of the acoustic scene. For these experiments four datasets of 10,000 synthetic binaural signals were generated using various signal types and processed using CEPBIMO. The binaural activity maps produced by CEPBIMO were evaluated using a series of simple convolutional neural networks. Though CEPBIMO was tested with more complex environments than preceding models its results exhibited more accuracy than results produced by other models and evaluated using more constrained environments.Ph
Intreactions between apolipoprotein e, tau protein, and 3-o-sulfated heparan sulfate: implications for alzheimer's disease pathogenesis
No full textJanuary 2024School of ScienceAlzheimer’s Disease (AD) is a neurodegenerative disease characterized by progressive and irreversible decline in memory and cognitive function. It is the leading cause of dementia and the 7th most prevalent cause of death in the United States. On a pathological basis, AD is characterized by the accumulation of proteinaceous inclusions within and surrounding the cells of the brain: plaques composed of amyloid β peptide, and neurofibrillary tangles (NFTs) within neurons composed of microtubule binding protein tau.Of these two hallmarks, the extent of tau NFT pathology is more predictive of AD progression and dementia symptoms compared to amyloid β. Tau pathology propagates in a prion like mechanism, whereby misfolded tau seeds propagate between cells and induce misfolding of endogenous tau protein. This process is mediated by an uptake pathway involving 3-O-sulfated (3-O-S) cell surface heparan sulfate proteoglycans (HSPGs) and the LRP1 receptor.
Intriguingly, HSPGs and LRP1 are also known to be utilized by Apolipoprotein E (ApoE), an important lipid transport protein whose gene (APOE) is the most significant risk factor for late onset Alzheimer’s Disease. In these studies, we present the results of investigations into the heparan sulfate (HS) binding behavior of ApoE and its competition with tau protein, underlying the important role of 3-O-S in Alzheimer’s Disease.
Utilizing a variety of biophysics and cell-based techniques, we demonstrated that ApoE binds preferentially to 3-O-S heparan sulfate, an interaction it shares with tau. Going further, we show via competition Surface Plasmon Resonance (SPR), Nuclear Magnetic Resonance (NMR) titrations, and Sedimentation Velocity Ultracentrifugation (SV-AUC) that ApoE and tau compete for binding to heparin and HS rather than forming a ternary complex.
Taken together with other recent developments in our understanding of tau pathology and its interplay with ApoE, this dissertation’s findings have significant implications for the role of ApoE in influencing AD pathogenesis and point to potentially novel avenues of AD drug and biomarker development.Ph
A batch technique for connecting multimodal ligand chemistry to chromatographic separability
No full textAugust 2024School of ScienceThis dissertation describes the development and utilization of a high-throughput parallel batch adsorption screen with sequential salt step increases to rapidly generate protein elution profiles for multiple resins at different pHs using a protein library. The chromatographic ligand libraries screened using this technique include commercial resins, Bio-Rad prototype small molecule resins, custom synthesized peptide-based ligands, and Solventum prototype membrane adsorbers. The chromatographic sets used in this work includes single, multimodal anion-exchange (MMA), and multimodal cation-exchange resins (MMC). The protein library consists of proteins with isoelectric points ranging from 3.4-11.4 with varying hydrophobicities as determined by their retention on hydrophobic interaction chromatography. The batch sequential experiments are carried out using one protein at a time with a wide set of resins at multiple pH conditions, thus enabling simple microtiter plate detection. A mathematical formulation is then used to determine the first moment of the distributions from each chromatogram (sequential step elution) generated in the parallel batch experiments. Batch data first moments (expressed in salt concentration) were compared to results obtained from column linear salt gradient elution, and the techniques are shown to be consistent. In addition, first moment data was used to calculate one-resin separability scores, which are a measure of a resin’s ability, at a specified pH, to separate the entire set of proteins in the library from one another. Again, the results from the batch and column experiments were shown to be comparable. The first moment data sets were then employed to calculate the two-resin separability scores, which are a measure of the ability of two resins to synergistically separate the entire set of proteins in the library. Importantly, these results based on the two-resin separability performances derived from the batch and column experiments were again shown to be consistent. This analytical approach was applied to all the screened ligand libraries and connections between ligand chemistry/stereochemistry and chromatographic behavior were identified. This approach for rapidly screening large numbers of chromatographic resins and mobile phase conditions for their elution behavior may prove useful for enabling the rapid discovery of new chromatographic ligands and resins.Ph
A data dictionary based approach to semantic tabular mapping
No full textAugust 2024School of ScienceKnowledge graphs are an important technology that enables a wide variety of analytics and data visualizations across an enterprise. However, creating knowledge graphs or adding to an existing knowledge graph can be challenging because data is often stored in a semi-structured form within tables that do not capture the full context of the data. To fill the context gap many data publishers include a data dictionary that aims to capture the meaning of schema elements, typically with text descriptions. These descriptions are helpful for humans to understand the data for integration tasks but are challenging for machines to interpret. Previous work has focused on integrating tables into an existing knowledge graph using data-level alignments without the additional context provided by data dictionaries. While these data-level alignment algorithms have proven successful on synthetic datasets, they struggle to make accurate alignments on real-world datasets that exhibit complex structures. Humans overcome these issues by leveraging context information from data dictionaries to understand the groups and relationships among the entities within a table. Recently, data publishers have started using this metadata to create semantic data dictionaries (SDDs) that formally capture alignments between tabular data. These alignments allow data publishers to convert tabular data into Resource Description Framework (RDF) triples and create or integrate data into a knowledge graph. However, SDDs require authors to have domain knowledge and experience in ontology modeling, which creates a barrier to entry for users. In this thesis, our goal is to improve the field of data integration by exploring algorithms that leverage context information from data dictionaries to align complex tabular data to ontology classes and properties. To achieve this, we address three key research questions: Can algorithms effectively use context from data dictionaries to improve alignment on complex tables? Are alignment algorithms that leverage data dictionaries competitive with data-level alignment algorithms on simple tables? What type of data dictionary descriptions are well suited for alignment algorithms? For the first research question, we developed the Semantic Data Dictionary Generator (SDD-Gen), a tabular alignment algorithm that generates SDDs by leveraging context information from data dictionaries. We show the effectiveness of SDD-Gen by comparing the performance against the current state of the art on complex tables. For the second research question, we developed a methodology for generating representative artificial data dictionaries using large language models. We use this methodology to generate data dictionaries for a popular tabular alignment dataset and show that SDD-Gen is as effective as the data-level algorithms on simple tables. For the final research question, we developed an evaluation framework to determine the type of data dictionary description best suited for tabular alignment. We show that intensional descriptions that define the conditions needed to be a member of a column are most effective and improve the reusability of data dictionaries.Ph
Enhancing efficiency of alternative energy conversion systems: a study of vibration and ocean thermal energy harvesters
No full textDecember 2023In all energy conversion systems and applications, efficiency is of major significance to engineers. This study focuses on two alternative energy conversion systems to improve their conversion efficiency. Part of this study is on electrostatic micro electromechanical systems (MEMS) used to harvest ambient kinetic energy or vibrations for applications such as powering Internet of Things (IoTs) sensors, while the other part is focused on marine thermal energy harvesting for powering Uncrewed Underwater Vehicles (UUVs).Electrostatic MEMS are used to harvest vibration energy from ambient sources such as wind, air ducts, human motion, and machinery. A typical electrostatic MEMS device consists of a spring-supported mass with a mobile electrode. This interacts with fixed electrode on the substrate to form a variable capacitor, converting mechanical energy into electrical energy when subjected to external vibrations. Despite the multitude of designs available for MEMS that harvest vibration energy, their maximum theoretical energy conversion limit remains largely unknown, hindering optimization efforts. This stands in contrast to other energy conversion systems like heat engines, which have a well-established theoretical limit in the form of Carnot efficiency. The concept of entropy served as the foundational element for deriving Carnot efficiency. This thesis explores the concept of mixing entropy, derived from statistical energy analysis, as an initial step towards discovering the efficiency limit of electrostatic MEMS. Specifically, mixing entropy is used to gauge the energy variation in an ideal electromechanical system at resonance. Nondimensional governing equations of the non-dissipative subsystems and the related different energy terms are developed. Multiple cases related to different initial conditions of the system are investigated. The results show that the maximum mixing entropy generated by the system coincides with the maximum energy transfer between the mechanical and electrical subdomains. A maximum defined effectiveness value of 4.4% for the system under consideration was obtained. This foundational research has the potential to be extended to more complex systems that include damping terms relevant to real-world applications. Such an expansion would pave the way for determining the maximum energy conversion limits for MEMS vibration energy harvesters.
While the above approach focuses on laying down fundamental concepts, the second approach takes more of a practical route. Specifically, to increase the efficiency of electrostatic energy harvesters, an impact-based frequency-up conversion is studied. This technique uses a combination of electrodes’ impact along with the impact of a springless mass (microball) to a shuttle mass. Frequency up-conversion generates high-frequency vibrations from low-frequency excitation, typically through electrode impact that triggers free oscillation. The objective of this study is i) to understand the feasibility of combining two distinct power enhancement methods (electrode impact and springless mass); and ii) to understand the effect of ball sizes and materials. The results suggest that the two methods can be combined seamlessly to produce a system with enhanced performance. Furthermore, from testing different ball materials (tungsten carbide, zirconium dioxide, and silicon nitride) and sizes it was found that optimum combinations depend on the applied bias voltage, acceleration, or frequency conditions. The addition of microball increases the device bandwidth especially at low vibration peak-to-peak acceleration of 0.5 g (g is earth’s gravitational acceleration).
The second part of this thesis targets the optimization of a system to be used in conjunction with a phase change material or PCM-based thermal engine to harvest ocean thermal energy. The oceans, largely unexplored, offer numerous resources and a rich biodiversity but present a harsh environment for exploration. UUVs serve as cost-effective tools for resource mapping, exploration, and scientific research. However, their mission durations are limited by battery life. To address this, researchers are turning to the ocean’s thermal energy as an alternative power source. Specifically, PCM-based systems can harness the ocean’s temperature gradient to generate power. As a UUV traverses different ocean depths, the PCM expands when absorbing heat from warmer surface water and contracts when releasing it in colder deep water. This volume change can pressurize a working fluid, which can either be used for propulsion through buoyancy changes or converted to electricity through a generator. While this offers a sustainable way to power the vehicle’s sensors and propulsion systems, the process’s efficiency is hampered by multiple stages of energy conversion. The goal of this thesis is to optimize this multi-stage conversion process in a real-world system. This exercise is first attempted on an ocean thermal energy using PCM to produce at least 8.1 kJ sufficient to power the SOLO-II float of the Argo program. A system-level numerical study is carried out simulating all key components. The optimized system is predicted to need less than 6 kg of PCM to produce the required power. The last part of the thesis expands on this type of system, investigating theoretically and experimentally a benchtop hydraulic-to-electric system for use with a PCM-based thermal engine. A numerical theoretical model is produced to simulate the performance of the hydraulic-to-electric system. Additionally, a predictive model using machine learning, specifically, the artificial neural networks (ANN), is developed for rapid assessment of the system energy based on available pressure and electrical load in use. Both the theoretical and ANN models are validated experimentally. The ANN model can monitor onboard overall energy conversion efficiency at a low computational energy demand with less than 15% relative error, allowing for better mission planning in real world UUV deployment. The maximum energy conversion efficiency of the benchtop model is 51%. The ANN model also has the potential for future expansion to include more system variables, paving the way for system-level energy efficiency optimization.Ph
Made of reclaimed fibers: an affordable, healthy and sustainable architectural material
No full textDecember 2023(Problem) The construction industry has faced the pressing challenge of reducing carbon emissions and improving health outcomes within the built environment. Conventional building materials contribute significantly to carbon emissions, and their production and disposal processes often pose health and environmental risks. Buildings and construction account for 39% of energy and process-related carbon dioxide emissions, 11% of which stems from the manufacturing of building materials and products. These staggering figures highlight the pressing need to shift towards sustainable building practices in order to mitigate the industry's adverse impact on the environment. To address these concerns, there is a requirement to explore alternative materials that are both sustainable and favorable to human well-being. (Hypothesis) One promising solution lies in utilizing agricultural and cellulosic fiber waste, combined with bio-binders, to develop biocomposite building materials for architectural interiors. Firstly, the use of such materials can lead to reduced carbon emissions in the life cycle of buildings, mitigating the industry's impact on climate change. Secondly, biocomposite materials have the potential to improve indoor air quality, as they do not release as harmful volatile organic compounds (VOCs) commonly found in traditional materials. This improvement in air quality can have a direct positive impact on the health and well-being of building occupants; as VOCs can have adverse effects on occupant health. Lastly, adopting such sustainable materials can promote economic responsibility by utilizing waste streams and reducing the reliance on resource-intensive manufacturing processes. (Methodology) The methodology employed in this research involves an exploration of bio-binders and different sources of cellulosic fiber waste, including agricultural waste, post-consumer waste, and industrial byproducts before delving into experimentation. The mechanical properties of various fiber types and combinations, as well as different binder and additive formulations, will be evaluated through flexural testing to identify the most viable recipes optimal for strong and stiff materials. Additionally, a comparison of the embodied carbon and other environmental impacts of the new materials with industry-standard materials are conducted to showcase the potential of fiber waste in driving sustainability and health in the building materials industry. (Impact) By investigating the viability and performance of biocomposite building materials derived from agricultural and cellulosic fiber waste, this research aims to contribute to the growing body of knowledge on sustainable construction practices. The findings will provide insights into the potential of these materials to lower carbon emissions, enhance indoor air quality, and promote economic responsibility within the built environment. Ultimately, by creating a new material that is functional, sustainable, and aesthetically competitive, this research endeavors to inspire designers, architects, and builders to embrace the potential of fiber waste and other underutilized resources in creating healthy and sustainable built environments.M