Indian Institute of Science Bangalore
etd@IISc Electronic Theses and Dissertations at Indian Institute of ScienceNot a member yet
6204 research outputs found
Sort by
Leveraging KG Embeddings for Knowledge Graph Question Answering
No full textKnowledge graphs (KG) are multi-relational graphs consisting of entities as nodes and relations
among them as typed edges. The goal of knowledge graph question answering (KGQA) is to
answer natural language queries posed over the KG. These could be simple factoid questions
such as “What is the currency of USA? ” or it could be a more complex query such as “Who
was the president of USA after World War II? ”. Multiple systems have been proposed in the
literature to perform KGQA, include question decomposition, semantic parsing and even graph
neural network-based methods.
In a separate line of research, KG embedding methods (KGEs) have been proposed to
embed the entities and relations in the KG in low-dimensional vector space. These methods
aim to learn representations that can be then utilized by various scoring functions to predict the
plausibility of triples (facts) in the KG. Applications of KG embeddings include link prediction
and KG completion. Such KG embedding methods, even though highly relevant, have not been
explored for KGQA so far.
In this work, we focus on 2 aspects of KGQA: (i) Temporal reasoning, and (ii) KG incompleteness. Here, we leverage recent advances in KG embeddings to improve model reasoning in
the temporal domain, as well as use the robustness of embeddings to KG sparsity to improve
incomplete KG question answering performance. We do this through the following contributions:
Improving Multi-Hop KGQA using KG Embeddings
We first tackle a subset of KGQA queries – multi-hop KGQA. We propose EmbedKGQA, a
method which uses ComplEx embeddings and scoring function to answer these queries. We find
that EmbedKGQA is particularly effective at KGQA over sparse KGs, while it also relaxes the
requirement of answer selection from a pre-specified local neighborhood, an undesirable constraint imposed by GNN-based for this task. Experiments show that EmbedKGQA is superior
to several GNN-based methods on incomplete KGs across a variety of dataset scales.
Question Answering over Temporal Knowledge Graphs We then extend our method to temporal knowledge graphs (TKG), where each edge in the KG
is accompanied by a time scope (i.e. start and end times). Here, instead of KGEs, we make
use of temporal KGEs (TKGE) to enable the model to make use of these time annotations and
perform temporal reasoning. We also propose a new dataset - CronQuestions - which is one of
the largest publicly available temporal KGQA dataset with over 400k template-based temporal
reasoning questions. Through extensive experiments we show the superiority of our method,
CronKGQA, over several language-model baselines on the challenging task of temporal KGQA
on CronQuestions.
Sequence-to-Sequence Knowledge Graph Completion and Question Answering
So far, integrating KGE into the KGQA pipeline had required separate training of the KGE
and KGQA modules. In this work, we show that an off-the-shelf encoder-decoder Transformer
model can serve as a scalable and versatile KGE model obtaining state-of-the-art results for
KG link prediction and incomplete KG question answering. We achieve this by posing KG link
prediction as a sequence-to-sequence task and exchange the triple scoring approach taken by
prior KGE methods with autoregressive decoding. Such a simple but powerful method reduces
the model size up to 98% compared to conventional KGE models while keeping inference time
tractable. It also allows us to answer a variety of KGQA queries, not being restricted by query
type
Quantum oscillation in band insulators and properties of non-equilibrium steady states in disordered insulators
No full textStarting with the experiment on Kondo insulator SmB6 , which shows 1/B-periodic oscillations despite the absence of gapless electronic excitations in bulk, the candidate insulators showing quantum oscillation (QO) are on the rise. But this is contrary to our conventional understanding that we need a Fermi surface to have QO. So an obvious question to ask is, ‘How can insulators show QO?’. If there is QO, then which physical quantities show QO, and what is the physical reason behind their origin? In the absence of a Fermi surface, what determines the frequency of these QO? In search of answers, we revisit recently proposed theories for this phenomenon, focusing on a minimal model of an insulator with a hybridization gap between two opposite-parity light and heavy mass bands with an inverted band structure. We show that there are characteristic differences between the QO frequencies in the magnetization and the low-energy density of states (LE-DOS) of these insulators, in marked contrast with metals where all observables exhibit oscillations at the same frequency. The temperature dependence of the amplitudes of the magnetization and DOS oscillations are also qualitatively different and show marked deviations from the Lifshitz-Kosevich form well-known in metals.
The interplay of disorder and interactions in quantum systems can lead to several intriguing phenomena, amongst which many-body localization (MBL) has caught many physicists’ attention in recent times. In the second work, we investigate whether an MBL system undergoes a transition to a current-carrying non-equilibrium steady state under a drive and how the entanglement properties of the quantum states change across the transition. The drive is introduced by using a phenomenological non-Hermitian model. We also discuss the dynamics, entanglement growth, and long-time fate of a generic initial state under an appropriate time evolution of the system governed by the non-H ermitian Hamiltonian. Our study reveals rich entanglement structures of the eigenstates of the non-Hermitian Hamiltonian. We find the transition between current-carrying states with volume-law to area-law entanglement entropy as a function of disorder and the strength of the non-Hermitian term.
In the third work, we take two 1D systems, namely the 1D Anderson model and Aubry-André-Harper model, that show the localization of single particle eigenstates depending on the strength of disorder and study it under a chemical potential drive. The drive is induced by connecting baths at different chemical potentials to the two edges of the system. We calculate the Green functions of the systems on the Keldysh contour that we use to calculate physical quantities like current and occupation of the non-equilibrium steady state. Our results can distinguish between the localized and delocalized phases in the non-interacting limit. In the presence of interaction, our systems can show MBL to a thermal phase transition. We end with a discussion on the possibility of probing this thermal to MBL phase boundary using dynamical mean field theory
A Monolithic Finite Element Formulation for Magnetohydrodynamics Involving a Compressible Fluid
No full textThis work develops a new monolithic finite-element-based strategy for Magnetohydrodynamics (MHD)
involving a compressible fluid based on a continuous velocity-pressure formulation. The entire formula-
tion is within a nodal finite element framework, and is directly in terms of physical variables. The exact
linearization of the variational formulation ensures a quadratic rate of convergence in the vicinity of
the solution. Both steady-state and transient formulations are presented for two- and three-dimensional
flows. Several benchmark problems are presented, and comparisons are carried out against analytical
solutions, experimental data, or against other numerical schemes for MHD. We show a good coarse-mesh
accuracy and robustness of the proposed strategy, even at high Hartmann numbers
Towards Robustness of Neural Legal Judgement System
No full textLegal Judgment Prediction (LJP) implements Natural Language Processing (NLP) techniques
to predict judgment results based on fact description. It can play a vital role as a legal assistant
and benefit legal practitioners and regular citizens. Recently, the rapid advances in transformer-
based pre-trained language models led to considerable improvement in this area. However,
empirical results show that existing LJP systems are not robust to adversaries and noise. Also,
they cannot handle large-length legal documents. In this work, we explore the robustness and
efficiency of LJP systems even in a low data regime.
In the first part, we empirically verify that existing state-of-the-art LJP systems are not robust.
We further provide our novel architecture for LJP tasks which can handle extensive text lengths
and adversarial examples. Our model performs better than state-of-the-art models, even in the
presence of adversarial examples of the legal domain.
In the second part, we investigate the approach for the LJP system in a low data regime. We
further divide our second work into two scenarios depending on the number of unseen classes in
the dataset which is being used for the LJP system. In the first scenario, we propose a few-shot
approach with only two labels for the Judgement prediction task. In the second scenario, we
propose an approach where we have an excessive number of labels for judgment prediction. For
both approaches, we provide novel architectures using few-shot learning that are also robust to
adversaries.
We conducted extensive experiments on American, European, and Indian legal datasets in the
few-shot scenario. Though trained using the few-shot approach, our models perform comparably
to state-of-the-art models that are trained using large datasets in the legal domain
Understanding the interplay between immune response and virus evolution
No full textRNA viruses are the underlying cause of many human diseases, including the common cold, influenza, dengue fever, and COVID-19. Their ability to evolve rapidly makes them challenging to tackle, resulting in public health threats. A well-known driver of virus evolution is the immune selection pressure, yet for many viruses, how it affects viral evolution remains poorly understood. How RNA viruses interact and evolve under the influence of the human immune system can help us develop effective treatments and better vaccines. At the same time, studying the immune response to viruses and their antigens can provide insights into how the human body develops immunity and how it can be boosted to fight infections.
In the first part of the thesis, we studied the diversity and evolution of the dengue virus (DENV) in India. Dengue is a mosquito-borne disease with four closely related virus serotypes (DENV1-4). About one-third of the global dengue cases are estimated to be from India. Yet we have a limited understanding of the dengue virus diversity and evolution in the country. Further, cross-reacting dengue antibodies from a prior infection from one serotype can protect or enhance infection from other serotypes. This can force the emergence of new dengue variants that find ways to escape the immune action and/or take advantage of it. In endemic countries like India, high rates of previous dengue infection can, hence drive the evolution of dengue serotypes in complex ways.
We sequenced 118 whole genomes from dengue patients collected over 7 years from four major cities across India. We combined them with all dengue sequences available fromIndia and compared them with the global strains of dengue. We examined the spatio-temporal dynamics of India-specific genotypes, their evolutionary relationship with global and local dengue virus strains, interserotype dynamics and their divergence from the vaccine strains. Our analysis highlights the co-circulation of all DENV serotypes in India with cyclical outbreaks every 3-4 years.
In South India, where seropositivity is very high, the envelope (E) protein displays strong signatures of evolution under immune selection. Apart from drifting away from its ancestors and other contemporary serotypes, we find evidence for recurring interserotype convergence, suggesting selection via antibody-dependent enhancement. We identify the emergence of the highly divergent DENV4-Id lineage in South India, which has acquired half of its E gene mutations in the antigenic sites. Moreover, the DENV4-Id is drifting towards DENV1 and DENV3 clades, suggesting the role of cross-reactive antibodies in its evolution. Our study shows how high incidence and pre-existing population immunity might shape dengue virus evolution in India.
In the second part of my thesis, I examine how antibody response to the SARS-CoV-2 virus evolves after COVISHIELD vaccination/ natural infection. We developed in-house ELISA assays to quantify the antibody levels (IgG and IgA) against the spike trimer, RBD and nucleocapsid proteins. Both IgG and IgA antibodies reached peak levels within 15-21 days of infection. Similarly, we observed an increase in antibody titers within the first 14 days post-COVISHIELD vaccination. The antibody levels waned over three months but were boosted after the second dose of the vaccine. We observed very high antibody titers in the case of COVID-19 recovered individuals with the first vaccination shot. The exposure due to infection also induces the IgA response, which is not mounted by vaccination alone. We found that although the antibody levels are lower, single vaccination can help reduce disease severity during breakthrough infections
Neuronal complex bursts and network information transfer in the hippocampus are robust to biophysical heterogeneities
No full textBiological entities must adopt mechanisms to override the impact of external perturbations to achieve stability and robustness. A crucial feature of biological systems is that they exhibit several forms of heterogeneities spanning all scales of functional analysis. A central question on biological robustness is therefore its relationship to heterogeneities, specifically addressing details pertaining to whether biological heterogeneities promote or impede robustness. In this thesis, we chose the mammalian CA3 sub-region of the hippocampus to be the system of interest towards understanding the impact of the biophysical heterogeneities on the functional robustness across the cellular and network scales.
Heterogeneities at the cellular scale are associated with the intrinsic properties of the CA3 pyramidal neurons as well as with synaptic inputs. The overall goal here was to assess the robust emergence of neuronal intrinsic properties (input resistance, back-propagating action potential amplitude, bursting and spiking profiles) along with complex spike bursting (CSB) in the CA3 pyramidal neurons with respect to heterogeneities in their parametric and measurement spaces. We generated a heterogeneous population of 12,000 random morphologically and biophysically realistic CA3 pyramidal neurons spanning a broad spectrum of parameters. We found two functional sub-classes of intrinsic bursting and regular spiking neurons, with significant differences in the expression profiles of N-type calcium and calcium-activated potassium (SK) channels. By triggering CSBs in all valid models using a variety of protocols, we observed substantial heterogeneities in the CSB propensities across models and protocols. Employing the virtual knockout approach for 7 different ion channels and N-methyl-D-aspartate receptors individually, we noted that synergistic interactions between several intrinsic and synaptic components regulated the robust emergence of CSB in these neurons. Together, we demonstrate the expression of ion-channel degeneracy in the robust emergence of physiological properties of CA3 pyramidal neurons including CSB, despite pronounced heterogeneities in their intrinsic and synaptic components.
Heterogeneities at the network scale are associated with intrinsic and synaptic components, with synaptic heterogeneities spanning local connections as well as afferent inputs from other brain regions. In this part of the thesis, we assessed the impact of neural-circuit heterogeneities, balance between excitatory and inhibitory synaptic strengths, and trial-to-trial variability on the spatial tuning profiles and spatial information transfer in the CA3 recurrent network. We employed homogeneous and heterogenous networks and stimulated them with spatially modulated inputs and employed the stimulus-specific information (SSI) metric to quantify the spatial information transfer by the place cells in these networks. We observed notable heterogeneities in spatial information transfer across both homogeneous and heterogeneous networks, with information transfer also dependent on synaptic inhibition strengths and trial-to-trial variabilities. Strikingly, spatial information transfer was robust to relatively higher noise levels in the heterogeneous networks compared to their homogeneous counterparts, thereby highlighting a crucial role for neural heterogeneities in enhancing the robustness of spatial information transfer in a recurrent place-cell network. We also found that a precise balance between recurrent and afferent connectivity was essential to maintain optimal spatial information transfer in neurons of such networks. Our analyses postulate a critical role for intrinsic heterogeneities in enhancing the robustness of spatial information transfer in a recurrent network of spatially tuned neurons.
Together, these analyzes point to a beneficial role for neural heterogeneities in the robustness of single-neuron and network physiology in the CA3 sub-region of the hippocampus
Establishment of a knock-in mouse model expressing a hypomorphic variant of receptor guanylyl cyclase C
No full textReceptor guanylyl cyclase C (GC-C, gene GUCY2C) is expressed on the apical surface of the intestinal epithelial cells and gets activated by the endogenous ligands guanylin and uroguanylin. Heat-stable enterotoxin (ST) secreted by enterotoxigenic E. coli is a super-agonist of GC-C. Activated GC-C catalyses the conversion of guanosine 5’- triphosphate (GTP) to cyclic guanosine 3’,5’-monophosphate (cGMP). GC-C via cGMP plays a key role in several biological processes such as the maintenance of intestinal fluid-ion homeostasis, regulation of intestinal cell cytostasis and tumorigenesis, mediation of gastrointestinal inflammation and protection against enteric pathogens like Salmonella Typhimurium.
Several disease-causing mutations have been identified in the GUCY2C gene in humans. Gain-of-function mutations result in increased GC-C activity leading to chronic diarrhoea and intestinal inflammation and the loss-of-function mutations result in meconium ileus due to the inactivation of the receptor. This study addresses the physiological implications of reduced cGMP by the generation of a novel knock-in mouse model harbouring a hypomorphic mutation in the linker region of GC-C. Mice with the hypomorphic mutant of GC-C showed reduced GC-C binding to ST indicating misfolding of the receptors. Further, the mice displayed decreased levels of cGMP in the colonic epithelial cells. Initial phenotypic characterisation of the mice showed no difference in their feeding behaviour, gut morphology and histology. However, a faster total gut transit and higher colonic inflammation were observed in mutant mice suggesting subtle changes occurring at the molecular level due to reduced cGMP.
To further explore the effects of decreased cGMP on the regulation of GC-C mediated protection against enteric pathogens, mice were infected with Citrobacter rodentium (C.r), a murine model for enteropathogenic E. coli infections. Using a bioluminescent tagged strain of C.r and live imaging of animals, higher pathogen load in faeces and higher bioluminescence was observed in the gut, suggesting increased colonisation of the pathogen and/or a more oxidative environment in the gut of mutant mice. Increased tissue-associated bacterial burden as compared to the wild type mice was observed in mutant mice resulting in higher C.r induced colonic inflammation. Reduced expression of C.r virulence factors with respect to the total pathogen colonisation was observed indicating the presence of avirulent C. rodentium in the intestine of knock-in mice, which could be attributed to the microbial dysbiosis.
Taken together, the novel mouse model developed in this study mimics humans harbouring heterozygous loss-of-function mutants in the GUCY2C gene that do not cause visible phenotype but would result in low cGMP levels in the gut. Furthermore, the model provides novel insights into the role of cGMP in the mediation of protection against enteric infections. Therefore, the knock-in mice can be used as a pre-clinical model to understand the consequences of low cGMP levels on gut physiology and to develop potential therapeutic strategies for alleviation of the symptoms associated with GC-C/cGMP mediated diseases
Global gyrokinetic simulations of electrostatic microturbulent transport in LHD stellarator and ADITYA-U tokamak
No full textTokamak and stellarator are two leading contenders in the quest to achieve nuclear fusion from magnetically-confined plasmas. They differ in terms of magnetic field structure in the toroidal direction. Irrespective of the magnetic field configuration, both the tokamak and the stellarator are prone to microturbulence, which is believed to be the major cause of particle and heat loss from the device. Thus, their understanding and control are paramount for the viability of nuclear fusion.
In this thesis, first-principles-based global gyrokinetic simulation studies of the electrostatic microturbulence are presented in the ADITYA-U tokamak, and Large Helical Device (LHD) stellarator and the effects of impurities on the microturbulence are investigated in both machines.
In the first part of the thesis, the global gyrokinetic simulations of the ion temperature gradient (ITG) and trapped electron mode (TEM) in the LHD stellarator are carried out with kinetic electrons using the numerically generated monotonic smooth plasma profiles. ITG simulations show that kinetic electron effects increase the growth rate and turbulent transport levels compared with simulations using adiabatic electrons. Zonal flow dominates the saturation mechanism in the ITG turbulence. However, its effect is weak on TEM turbulence. Following this, a realistic experimental discharge is analyzed in the presence of boron impurities. Simulations show the co-existence of ITG and TEM turbulence, with the linear frequencies matching well with the experimental observations. Nonlinear simulations show that the reduction in nonlinear transport is a combined effect of the change in plasma profile and plasma dilution due to boron impurities.
In the second part of the thesis, the global gyrokinetic simulations of the electrostatic microturbulence driven by the pressure gradients of thermal ions and electrons are carried out for the ADITYA-U tokamak geometry using its experimental plasma profiles and with collisional effects. The dominant instability is TEM, based on the linear eigenmode structure and its propagation in the electron diamagnetic direction. Collisional effects suppress turbulence and transport to a certain extent. Simulations by artificially suppressing the zonal flow show that the zonal flow is not playing a critical role in the TEM saturation, which is dominated by the inverse cascade. The frequency spectrum of the electrostatic fluctuations is in broad agreement with the experimentally recorded spectrum. Following this, the effect of argon impurities on turbulence and transport is investigated. Simulations show that the primary mechanism responsible for the reduction in transport is the change in plasma profile due to argon puffing.
Finally, a novel framework is presented in cylindrical coordinates to get rid of the difficulties of the null point (X-point), where the poloidal magnetic field vanishes, along with the singular behavior of the safety factor and Jacobian in Boozer coordinates that enables the whole volume gyrokinetic simulations of fusion plasma.IPR, UCI, NSM, SERB, BRNS, Infosys, DAE India, DoE U
Index Coding over Noisy Channels and Some Applications
No full textA broadcast channel that is very effective for disseminating common content becomes highly inefficient when the users request different content. To address this inefficiency of the broadcast channel over which a server transmits distinct contents to different caching receivers, the concept of index coding was introduced by Birk and Kol in ``Informed-source coding-on-demand (ISCOD) over broadcast channels'' in Proceedings. IEEE INFOCOM, 1998. In an index coding problem (ICP), a server with access to a set of messages broadcasts over a forward channel to a group of caching receivers. Each receiver has a subset of the messages available in its cache and requests a non-intersecting subset of messages from the server. The server is informed about the cached contents at the receivers through a slow backward channel. Index coding aims to satisfy the message requests of all the receivers with the minimum number of server transmissions by utilizing the information of receivers' cached contents and their data requests. This problem, a variant of source coding with side information, has applications in several engineering problems in network communication, such as content broadcasting, device-to-device communication, distributed caching, distributed computation, and interference management. While most of the literature on index coding focuses on noiseless broadcast, our work focuses on various sub-classes of index coding problems and their applications, predominantly when the transmissions are over noisy broadcast channels.
We consider a binary-modulated transmission of index codes over continuous-output channels for single uniprior ICPs and develop algorithmic solutions which have reduced average probability of error. For the noisy ICP with M-ary modulated transmission, the ML decoding strategy proposed by Sudhakaran et al. in ``Index Coded PSK Modulation for Prioritized Receivers," in IEEE Transactions on Vehicular Technology, Dec. 2017, resulted in the question of whether the central server's encoding scheme should be an index code or whether any set of transmissions encoding across the entire library of messages is sufficient. We prove that index codes are necessary for solving such problems. We also address how to map index-coded vectors to signal points on a square QAM constellation and prove that unlike in uncoded point-to-point communication, cases can arise in a noisy ICP setting where M-PSK outperforms M-QAM. For the same class of ICPs, we propose multi-symbol PSK-modulated transmission, where the index-coded bits are mapped to complex symbols of multiple smaller-sized constellations for a bandwidth-performance trade-off.
Error correction for index coding was first considered by Dau et al. in the paper ``Error correction for index coding with side information,'' in IEEE Transactions on Information Theory, March 2013. A delta-error correcting index code is an encoding of the messages such that all the receivers can decode their required messages correctly in the presence of at most delta errors. Concatenating an optimal linear index code and a classical error-correcting code of minimum length, which can correct delta errors, will not always lead to optimal delta-error-correcting index codes. We develop optimal linear error-correcting transmission schemes for three different settings: a multi-access coded caching setting, a device-to-device coded caching setting, and a coded distributed computing setting. A few results in embedded index coding and coded caching derived using index coding techniques are also discussed. Finally, for two distributed computing settings, we propose a technique to reduce the recovery threshold, which is the minimum number of workers who should return their results for the master to be able to complete its task.
In short, this thesis focuses on various sub-classes of index coding problems and their applications in device-to-device communication, coded caching, and distributed computing, predominantly when the transmissions are over noisy broadcast channels
Human Aware Path planning for AMRs in dynamic Warehouse Environment
No full textWith the increasing integration of robots into various domains and their coexistence with humans, ensuring human safety has become a critical concern. This thesis highlights the challenges associated with human safety in human-robot coexistence and provides an overview of the solutions and approaches proposed to address these challenges. The coexistence of humans and robots introduces unique safety considerations due to human behavior's dynamic and unpredictable nature and the potential risks posed by robotic systems. Risk assessment involves several steps, including identifying potential hazards, risk analysis, and implementation of safety measures. Safety standards are crucial in ensuring robots' safe and responsible deployment in human-centric environments. Safety standards prioritize human well-being by reducing the risk of physical harm, ensuring safe operation, and establishing guidelines for safe design and behavior. Compliance with these standards enhances public trust in robotic systems, encouraging widespread acceptance and adoption. This thesis explores the standards organizations and regulatory bodies associated with robotics and related industries. Some prominent organizations include the International Organization for Standardization, the International Electrotechnical Commission (IEC), the American National Standards Institute (ANSI), and industry-specific associations like the Robotic Industries Association (RIA).
Human trajectory prediction anticipates individuals' positions and movements in dynamic environments. Data-driven methods can be applied to predict the human trajectory. These methods leverage historical trajectory data, environmental context, and other relevant features to learn patterns and make predictions. The thesis work aims to make the robots navigate in a collaborative environment without invading human space. Robot navigation duties must be carried out fast and safely in the presence of humans, which necessitates making predictions about their intended trajectories. This work explores the usage of probabilistic models relating to human behavior perceived from the environment for the motion planning of Autonomous Mobile Robots. We study the use of a) Maximum-likelihood Human Trajectory Prediction and b) Gibbs-sampled Human Trajectory Prediction. The forecasted human trajectory is used to determine a safe path for robots. We study the efficacy of including the human motion predicted during the motion planning in terms of the effective speed of the robots, with and without human-aware planning