DSpace@RPI (Rensselaer Polytechnic Institute)
Not a member yet
6809 research outputs found
Sort by
Incorporating external knowledge into food representation and recommendation
No full textAugust 2022School of ScienceFood is our most basic need, the very stuff of life. Food is fundamental for human beings and eating well is essential to good health.
Food computing has attracted great interest from various fields.
Representing food as vectors can capture hidden information gleaned from massive datasets and help further studies on food.
As an important task of food computing, food recommendation can be a means to help people find food they might love,
and also to help them eat more healthily.
Acting as a special case of food recommendation, food substitution is an open task to explore potential food alternatives.
Structured knowledge furnishes an in-depth understanding of the world.
At the same time, knowledge graphs (KG) that represent structural relations between entities have become an increasingly popular research tool towards enabling artificial intelligence.
KGs can boost the studies in food computing by its structural and semantic information as well as providing interpretability.
In this thesis, we focus on two main research questions:
(1) How can we capture the semantics of recipes for use in food representation and recommendation (QR1)?
(2) How can we employ external knowledge like KGs to facilitate tasks of food computing (QR2)? To address the two research questions, we propose a novel model for recipe representation learning from pure text.Learning recipe embeddings is a challenging task, since there is a lack of high quality annotated food datasets.
We tackle this problem by adopting a triplet loss for model optimization so that related recipes are closer in the latent semantic space.
An external knowledge source like the food KG is employed to construct feasible recipes triples by performing recipe sample mining.
We provide a joint approach for learning effective pretrained recipe embeddings using both the ingredients and cooking instructions.
A set transformer is adopted to encode the ingredient set to preserve its permutation-invariant property. In terms of food recommendation, it is challenging to offer users food that both meets users' preference and a health goal.We introduce health-guided recipe recommendation as a way to incrementally shift users towards healthier recipe options while respecting the preferences reflected in their historical choices.
The food KG aids this task by providing relationships among food along with their nutrition information.
Thus, we consider the task of recipe recommendation over KGs.
In particular, we jointly learn recipe representations via graph neural networks over two KG subgraphs, which target user preferences and recipe healthiness.
Another challenge in food recommendation is that it is hard to balance the trade-off between preference and healthiness.
We utilize a knowledge transfer scheme to enable the transfer of useful semantic information across the preferences and healthiness aspects instead of simple fusion. Combining the above two tasks, we further propose an approach to learn KG representations with adversarial food substitution.To enlarge the scope of food representation learning that is not limited to recipe data and to fully utilize the external knowledge, we perform knowledge representation learning over a food KG.
We employ a pretrained language model to encode entities and relations, thus emphasizing contextual information in food KGs.
The model is trained on two tasks -- predicting a masked entity from a given triple and predicting the plausibility of a triple obtained from the KG.
As an open food recommendation task, analysis of food substitutions helps in identifying optimal dietary choices and supporting different user needs.
It is hard to evaluate the substitutions due to the lack of an adequate validation set.
To tackle this challenge, we propose a collection of adversarial sample generation strategies for different food substitutions over our learnt KG embeddings.
To meet different purposes, we generate high quality context-aware recipe and ingredient substitutions by replacing, adding, and detecting actions over token and entities.
We also provide generalized ingredient substitutions to meet the needs of general substitution purpose.Ph
GRASP depletion-mediated Golgi fragmentation impairs glycosaminoglycan synthesis, sulfation, and secretion
No full textCellular and Molecular Life Sciences, 79, 199Note : if this item contains full text it may be a preprint, author manuscript, or a Gold OA copy that permits redistribution with a license such as CC BY. The final version is available through the publisher’s platform.Synthesis of glycosaminoglycans, such as heparan sulfate (HS) and chondroitin sulfate (CS), occurs in the lumen of the Golgi, but the relationship between Golgi structural integrity and glycosaminoglycan synthesis is not clear. In this study, we disrupted the Golgi structure by knocking out GRASP55 and GRASP65 and determined its effect on the synthesis, sulfation, and secretion of HS and CS. We found that GRASP depletion increased HS synthesis while decreasing CS synthesis in cells, altered HS and CS sulfation, and reduced both HS and CS secretion. Using proteomics, RNA-seq and biochemical approaches, we identified EXTL3, a key enzyme in the HS synthesis pathway, whose level is upregulated in GRASP knockout cells; while GalNAcT1, an essential CS synthesis enzyme, is robustly reduced. In addition, we found that GRASP depletion decreased HS sulfation via the reduction of PAPSS2, a bifunctional enzyme in HS sulfation. Our study provides the first evidence that Golgi structural defect may significantly alter the synthesis and secretion of glycosaminoglycans.National Science Foundatio
Reduced-order modeling of neutron transport by proper generalized decomposition
No full textAugust 2022School of EngineeringThe numerical simulation of neutron transport within a nuclear reactor—for brevity, reactor physics—is the foundation on which analysts: design new reactors; license existing designs; prolong the licenses of operating plants; optimize loading patterns of fuel and burnable absorbers; perform fuel cycle analyses; evaluate a reactor’s ability to self-regulate in response to perturbations; model reactivity control systems; design adequate shielding; and predict the isotopic composition of spent fuel. It is, in short, how humans understand the neutronic inner workings of a nuclear reactor, second only to physical experimentation. It is also, unfortunately, extremely resource-intensive to deterministically compute at high fidelities, being described as a six- or seven-dimensional integro-differential equation. These dimensions refer to the neutron field’s position in space r⃗ ≡ (x,y,z), direction of travel (in angular coordinates μ and ω), speed (energy) E, and instant in time t (for transients). As such, discretizing the neutron transport equation with even a scant ten unknowns in each dimension would yield a seven-dimensional mesh containing ten million degrees-of-freedom. Unsurprisingly, in practical simulations of large reactors, this number often runs into the billions or trillions, requiring either vast High Performance Computing (HPC) resources or drastic simplifications. This phenomenon is known generally as the “curse of dimensionality” and is common to many fields of numerical analysis. Presently, we aim to circumvent the curse of dimensionality by seeking a separable, low-rank approximation of the neutron flux, or solution. Moreover, unlike a posteriori, or data-driven, Reduced-Order Models (ROMs), this decomposition will be computed progressively by way of a greedy algorithm, eliminating the need for full-order reference solutions. Specifically, the a priori model order reduction technique here applied to reactor physics is Proper Generalized Decomposition (PGD). Because PGD approximates the solution to high-dimensional problems like radiation transport as a finite series of M products of low(er)-dimensional modes, one avoids solving the high-dimensional (full-order) problem entirely. Instead, only M nonlinear systems of low-dimensional subproblems need be solved; as such, the PGD ROM may be drastically cheaper to compute than the original problem, especially if few modes M are needed. Particularly, we here separate energy (yielding spatio-angular and energetic subproblems) and axial space (yielding 2D and 1D subproblems). Both ROMs are then validated in prototypical reactor physics benchmarks. Our first application—model order reduction in energy by PGD—is motivated by the extreme (ultrafine) resolution required to resolve individual nuclear resonances—sharp peaks across a narrow range of energy—in the neutron interaction probabilities, or cross sections, of nuclear fuels and other materials. This regime of fidelity is so demanding as to be typically impractical for geometries larger than a 1D or 2D slice across a single fuel pin—let alone the hundreds of pins comprising an assembly (or the hundreds of assemblies comprising a core). For now, however, we consider the more modest energy meshes (70 to 361 groups) usually employed for infinite lattices of pins or assemblies (that is, lattice physics), such that it is tractable to compute the full-order solution for comparison. Benchmark cases are taken to be representative light water reactor (LWR) pins of UO2 or Mixed Oxide fuel with CASMO-70, XMAS-172, and SHEM-361 energy meshes. To begin, we establish that both the Galerkin and Minimax PGD ROMs are able to compute the flux at sufficient precision (0.36% L2 error or less) in a tractable number of modes (M = 50). Next, we apply the ROM to cross section generation—often the objective of lattice physics—achieving results comparable to a homogenized, infinite medium model with 1 to 3 modes and comparable to the full-order model by 10 to 20 modes, as assessed by the error of the coarse-group model. Subsequently, we compare the coarse(ned)-group flux against that given by cross section condensation, finding similar L2 errors (0.5%) with 10 modes. Given additional modes, the ROM is able to converge below this threshold. Finally, this ROM is extended to criticality (eigen)problems by means of an original algorithm, which achieves k-eigenvalue errors less than 2 × 10−4 by M = 50. Further, the eigenvalue ROM again compares favorably to the coarse-group model with as few as 10 to 20 modes. Based on these results, we anticipate this PGD ROM may be able to calculate detailed flux distributions and cross sections more economically than the full-order model, at a marginal or negligible detriment to accuracy. Moreover, the ROM presents an alternative means of approximation to cross section condensation, preferable in that it introduces neither a loss of fidelity nor irrecoverable error. Secondly, we apply PGD to separate the axial and (optionally) polar dimensions of neutron transport. As nuclear reactors (especially LWRs) tend to be tall, but geometrically simple, in the axial, or z direction, we expect this ROM may save substantial effort and rapidly converge to a low-rank approximation. Moreover, we anticipate this approach may compare favorably to the methodologically distinct, but practically analogous 2D/1D methods already practiced in reactor physics. First, we derive two original models: that of axial PGD—which separates only z and the axial streaming direction v ∈ {−1,+1}—and axial-polar PGD—which separates both z and polar angle μ from the radial domain. Additionally, we grant that the energy dependence E may be ascribed to either radial or axial modes, or both, bringing the total number of candidate 2D/1D ROMs to six. To assess performance, these PGD ROMs are then applied to two few-group benchmarks characteristic of LWRs. Therein, we find each ROM to be convergent and the axial-polar PGD to be often more economical than the axial PGD. Ultimately, given the popularity of 2D/1D methods in reactor physics, we expect a PGD ROM which achieves a similar effect, but perhaps with superior accuracy, a quicker runtime, and/or broader applicability, would be eminently useful, especially for full-core problems. Finally, we discuss the neutron transport software developed to implement both the the full-order and PGD models, Aether. More specifically, in order to meaningfully apply these ROMs it was necessary to first establish a basic set of features—namely, unstructured mesh geometry, spatial discretization by finite elements, and hyperbolic transport with matrix-freesweeps. Since no software was available that met these requirements, we here develop an original, C++ library, in turn using the deal.II finite element package. Despite the specialized research objectives above, the software is organized such that the particularities of PGD do not appear in the full-order model, but rather are implemented as wrappers around or modifications of it. This allows the library to serve as a general-purpose research tool for deterministic radiation transport, even outside of applications in PGD. Ultimately, we intend to release Aether as a permissively open-source software library, such that others can use and modify this implementation at will. Moreover, while some outstanding features (preconditioning, parallelism) would be practically required, we envision with a modest effort, Aether could be made a useful application for end-users, not just developers, akin to OpenMC or OpenMOC.Ph
Coping with shock : numerical procedure for the detection and sorting of shocks in fluid flows
No full textMay 2021School of EngineeringWithin computational fluid dynamics (CFD) simulations, many complex flow phenomena can emerge. Accurately modeling these phenomena is often a primary motivating factor behind these simulations, and shocks are some of the most challenging phenomena to either model or resolve. However, once a shock is identified and well defined within a flow, there are many proven tools available to both improve the resolution of a given shock, and to efficiently propagate it through space. These methods often rely upon key geometric and/or mesh information related to the shocks, but such knowledge is non-trivial to obtain apriori. This thesis presents a numerical procedure to efficiently detect shocks, filter them to eliminate noise, and sort them into separate shocks or individual shock segments for further analysis. The resulting information could be used for shock fitting, anisotropic layered meshing, or analysis of complex shocks. A combination of numerical simulation data and manufactured data is used to demonstrate the performance of the current procedure. Overall, the novelprocedure developed in this work is robust, and results obtained thus far demonstrate its efficacy.M
Ion-conducting polymers via acid-catalyzed polyhydroxyalkylation for electrochemical devices
No full textDecember 2021School of ScienceThe dependency of modern-day human life on synthetic polymers is becoming increasingly evident. At present, polymers are used in a broad spectrum of applications, ranging from commodity polymers in everyday plastics, to functionalized polymers in highly specialized applications. This thesis focuses on one specific application, i.e., the use of ion conducting polymers as polymer electrolyte membranes in electrochemical energy conversion and storage systems, such as fuel cells. Among many polymerization techniques, the use of acid-catalyzed polyhydroxyalkylation reaction for the synthesis of polymers that are later functionalized into ion conducting polymers is highlighted herein. In the recent years, this reaction has gained significant popularity due to its synthetic convenience, capacity for scaled-up synthesis, and inexpensive raw materials. A major advantage of this polymerization method is the assortment of monomers that can be used to produce polymers with desired chemical structures. This thesis discusses three separate projects. The first two projects are aimed at exploring different monomers to create anion conducting polymers with unique structural diversities. In order to do so, variations were carried out at the polymer backbone structure, cation distribution and side chain length. The properties of these anion conducting polymers were studied in relation to their structure, and their suitability in respect of specific requirements of anion exchange membrane fuel cells are discussed. these studies reflect the ability of this reaction to create specialized ion conducting polymers by tailoring polymer properties depending on the desired function or purpose.
In the third project, the potential of a selected series of under-exploited monomers in acid-catalyzed polyhydroxyalkylation reaction were investigated. The studies were aimed at broadening the scope of the reaction and discovering environmentally friendly fluorine-free polymers. By converting the polymers into proton conducting membranes, the application of these polymers in high-temperature proton exchange membrane fuel cells was demonstrated. This study serves as a steppingstone to discovering novel polymer structures synthesized by acid-catalyzed polyhydroxyalkylation and the diverse utility of its products.Ph
Intrinsically disordered N-terminal domain (NTD) of p53 interacts with mitochondrial PTP regulator Cyclophilin D
No full textJournal of Molecular Biology, 434, 167552Note : if this item contains full text it may be a preprint, author manuscript, or a Gold OA copy that permits redistribution with a license such as CC BY. The final version is available through the publisher’s platform.Mitochondrial permeability transition pore (mPTP) plays crucial roles in cell death in a variety of diseases, including ischemia/reperfusion injury in heart attack and stroke, neurodegenerative conditions, and cancer. To date, cyclophilin D is the only confirmed component of mPTP. Under stress, p53 can translocate into mitochondria and interact with CypD, triggering necrosis and cell growth arrest. However, the molecular details of p53/CypD interaction are still poorly understood. Previously, several studies reported that p53 interacts with CypD through its DNA-binding domain (DBD). However, using surface plasmon resonance (SPR), we found that both NTD-DBD, NTD and NTD (1-70) bind to CypD at ∼μM KD. In solution NMR, NTD binds CypD with μM affinity and mimics the pattern of FLp53 binding in chemical shift perturbation. In contrast, neither solution NMR nor fluorescence anisotropy detected DBD binding to CypD. Thus, instead of DBD, NTD is the major CypD binding site on p53. NMR titration and MD simulation revealed that NTD binds CypD with broad and dynamic interfaces dominated by electrostatic interactions. NTD 20-70 was further identified as the minimal binding region for CypD interaction, and two NTD fragments, D1 (residues 22-44) and D2 (58-70), can each bind CypD with mM affinity. Our detailed biophysical characterization of the dynamic interface between NTD and CypD provides novel insights on the p53-dependent mPTP opening and drug discovery targeting NTD/CypD interface in diseases
The (Uncomputable!) Meaning of Ethically Charged Natural Language, for Robots, and Us, from Hypergraphical Inferential Semantics
No full textThe year is 2030. A two-young-child, two-parent household, the Rubensteins, owns and employs a state-of-the-art household robot, Rodney. With the parents out, the children ask Rodney to perform some action ��
that violates a Rubensteinian ethical principle ����
. Rodney replies: (��1
) “Doing that would be (morally) wrong, kids.” The argument the children give Rodney in protest is that another household, the Müllers, also has a robot, Ralph; and the kids argue that he routinely performs ��
. As a matter of fact, Ralph’s doing ��
violates no Müllerian ethical principle ����
. Ralph’s response to the very same request from the children he tends is: (��2
) “Okay, doing that is (morally) fine, kids.” What is the meaning of the utterances made by Rodney and Ralph? We answer this question by presenting and employing a novel, formal, inferential theory of meaning in natural language: hypergraphical inferential semantics (ℋℐ��
), which is in the general spirit of proof-theoretic semantics, which is in turn antithetical to Montagovian model-theoretic semantics. ℋℐ��
, applied even to sentences logically simpler than ��1
and ��2
, implies that human-level natural language understanding (NLU) is Turing-uncomputable.10.1007/978-3-031-09823-9_1
Learning in limited-data and limited-annotation scenarios
No full textMay 2022School of EngineeringDeep learning has demonstrated its capability in a wide range of learning tasks. However, it generally requires a vast amount of annotated data to train deep models. In many areas, creating such datasets consumes a considerable amount of resources, time, and effort. This dramatically restricts the applicability of deep learning in many real-world scenarios. It is thus of paramount importance to develop deep learning models that can leverage the small amount of annotated data available. There are many restricted domains where a limited supply of training data is available. In this situation, transfer learning and few-shot learning can help mitigate the limited data issue. In transfer learning, representations learned from a large dataset are re-purposed to another domain that has limited annotated data. The general approach in transfer learning is to use the ImageNet-pretrained ResNet model as a feature extractor, and either train a classifier header on top of that or finetune the full network to predict the classes of the downstream dataset. However, we show that ResNet models trained with self-supervised losses, particularly self-supervised contrastive losses, can have better transferability than those trained with cross-entropy loss. We study the transferability of learned representations of different contrastive approaches in downstream linear evaluation, full-network transfer, few-shot recognition, and object detection tasks. The results show that the contrastive approaches learn representations that are easily transferable to different downstream tasks. We further observe that a joint objective of self-supervised contrastive loss with cross-entropy/supervised-contrastive loss leads to better transferability of these models over their supervised counterparts. Most self-supervised learning models rely on image classification tasks for evaluation. Due to the global (i.e., single label) nature of image classification, the existing methods use a single feature representation of the image to guide the respective loss functions. Many computer vision tasks (e.g., object detection, semantic segmentation, pose estimation) require spatially localized feature representations. We present a novel self-supervised learning method that imposes local consistency between corresponding regions of transformed versions of the same image, which can be used alongside any self-supervised learning method with minimum computational overhead. We report significant improvement over existing self-supervised pre-training methods with this approach for object detection and semantic segmentation. Next, we tackle the problem of cross-domain few-shot learning where there is a large shift between a base dataset that has many labeled samples and a target domain which mostly has a few labeled examples and many unlabeled samples. The goal is to learn a feature extractor network from the labeled base dataset and then adapt the weights from the few labeled samples from the target dataset. We propose a simple dynamic distillation-based approach to use unlabeled images from the target dataset. We impose consistency regularization by calculating predictions from the weakly-augmented versions of the unlabeled images from a teacher network and matching it with the strongly augmented versions of the same images from a student network. We show that the proposed network learns a representation that can be easily adapted to the target domain even though it has not been trained with target-specific classes during the pretraining phase. One particular domain where annotation is expensive is video action localization. It is very time consuming to label all start and end frames of every action in an untrimmed video. We develop a weakly-supervised action localization model that is trained on videos with only video-level action categories without any temporal supervision, and can predict the start and ending frames for each action in the video. We developed two methods to solve the task. In the first approach, we use deep metric learning to learn the embedding of video segments such that videos with similar actions are closer in the embedding space and videos with different actions are farther away. We propose a classification module to generate action labels for each segment in the video, and a deep metric learning module to learn the similarity between different action instances. We jointly optimize a balanced binary cross-entropy loss and a metric loss using a standard backpropagation algorithm. Next, we present a hybrid attention mechanism for weakly supervised temporal localization. We argue that existing multiple instance learning (MIL) approaches have a major limitation of only capturing the most discriminative frames of an action, ignoring the full extent of an activity. Moreover, these methods cannot model background activity effectively, which plays an important role in localizing foreground activities. We develop a novel hybrid attention mechanism that includes temporal soft, semi-soft and hard attentions to address these issues.Ph
Numerical study for CH4 production from gas hydrate-bearing sediments via CO2 injection
No full textAugust 2021School of EngineeringCH4 production from gas hydrate-bearing sediments has been considered as a potential measure to supplement global hydrocarbon resources. Natural gas hydrates, solid compounds of gas and water, are often found in the pores of sediments under deepwater or permafrost regions, where hydrate stability conditions of high pressure and low temperature exist. CH4 production from gas hydrate-bearing sediments requires a phase change of the solid gas hydrates and has been deemed possible by depressurization through a well. However, the method tends to lower geomechanical stability of the sediments and consequently, to date, only a few short-term trials of field-scale gas production were successful. Moreover, future hydrocarbon production is overshadowed by growing concerns for its adverse environmental impacts. Therefore, there is a pressing need for an innovative energy strategy that replenishes dwindling supplies of hydrocarbons while maintaining geomechanical stability and achieving carbon neutrality. CO2 injection into gas hydrate-bearing sediments may potentially accomplish the aforementioned objective. As CO2 hydrates are usually more stable than CH4 hydrates, injected CO2 could destabilize CH4 hydrates in the sediments to release out CH4 gas for production while forming solid CO2 hydrates in the newly available pore of the host sediments. In this way, it would effectively maintain the geomechanical stability of the sediments during CH4 production and, at the same time, store CO2 as solid hydrates permanently in the sediments. However, the process is yet to be fully understood as it involves interactions of various multi-physical and chemical processes including generation of immiscible CH4-CO2 fluid mixture in sea water, evolution of chemical reaction kinetics of CH4-CO2 hydrate mixture, heat emission and absorption from hydrate formation and dissociation, respectively, and stress redistribution caused by spatial and temporal developments in CH4-CO2 mixed hydrate-bearing sediments. This research has developed a novel, coupled thermo-hydro-chemo-mechanical formulation that captures the complex processes and has investigated the behavior of CH4 hydrate-bearing sediments subjected to CO2 injection. There are mainly four contributions. Firstly, several additional processes caused by generation of CH4-CO2 fluid mixture have been incorporated such as fluid viscosity change and molar fraction induced diffusion. Secondly, formation of CH4-CO2 hydrate mixture has been taken into account, along with its effect on thermal, hydrological, chemical and mechanical processes. Thirdly, the applicability of the formulation has been validated through simulations of existing laboratory tests of CO2 injection into CH4 hydrate-bearing soil. Fourthly, the efficiency of CH4 production and CO2 storage and its geomechanical impact by CO2 injection in natural gas hydrate-bearing sediments have been thoroughly discussed. The outcome of this research provides valuable insights into the prospect of the revolutionary energy strategy that satisfies conflicting interests, unlocking the new source of hydrocarbon energy while managing geo-risk and environmental sustainability for future generations.Ph
Circadian Control of Heparan Sulfate Levels Times Phagocytosis of Amyloid Beta Aggregates
No full textPLOS Genetics, 18(2): e1009994Note : if this item contains full text it may be a preprint, author manuscript, or a Gold OA copy that permits redistribution with a license such as CC BY. The final version is available through the publisher’s platform.Alzheimer’s Disease (AD) is a neuroinflammatory disease characterized partly by the inability to clear, and subsequent build-up, of amyloid-beta (Aβ). AD has a bi-directional relationship with circadian disruption (CD) with sleep disturbances starting years before disease onset. However, the molecular mechanism underlying the relationship of CD and AD has not been elucidated. Myeloid-based phagocytosis, a key component in the metabolism of Aβ, is circadianly-regulated, presenting a potential link between CD and AD. In this work, we revealed that the phagocytosis of Aβ42 undergoes a daily circadian oscillation. We found the circadian timing of global heparan sulfate proteoglycan (HSPG) biosynthesis was the molecular timer for the clock-controlled phagocytosis of Aβ and that both HSPG binding and aggregation may play a role in this oscillation. These data highlight that circadian regulation in immune cells may play a role in the intricate relationship between the circadian clock and AD.National Science Foundatio