Abstract

The Large Hadron Collider (LHC) experiment is searching for exotic beyond the Standard Model signal processes. These processes occur at extremely low rates compared to the Standard Model ones making it difficult to extract these signals effectively. Classical cut-based methods that rely heavily on hand-crafted strategies fail to perform well on these exotic signals. For the next stage of runs at the LHC, there is an urgent need to develop newer algorithms more capable of identifying these rare signals. Deep learning methods have recently gained traction to assist these searches due to their exceptional ability to analyse complex patterns and show immense potential to improve the sensitivity of experiments. In this thesis, I describe a few approaches to adapt and incorporate deep learning techniques into collider experiments. First, I present a case study of a phenomenological search for a hypothetical heavy quark, where I augment the standard search strategy with a simple deep-learning model to yield better sensitivity and reach. In the case study, I also discuss the caveats of such methods and point to a potential solution. Secondly, I present a simple adaptation of the deep learning training strategy to align it with physics goals. Specifically, I study modifications to the standard Cross Entropy loss and propose a new loss function, based on the Lovasz extension, to directly optimise the sensitivity metric (Z-score). Finally, I present a novel generative model based on the GAN (Generative Adversarial Framework) to generate particle jets. This model can quickly generate jets from different processes and assist jet studies.

Abstract

Metaverse is predicted to be the next game changer after artificial intelligence. In this work, we operationally define metaverse applications as virtual reality (VR) applications with possibly multi-modal, multi-user, and multi-sensory inputs. The idea of creating a VR space with a hardware-software co-design system can enhance user interaction. Additionally, the VR based systems interacting in the 3-dimensional environment can be beneficial in addressing the socio-economic reforms currently faced in society. The potential extent that VR systems can be used in medical, educational and defence capabilities is substantial. Such scale can be recorded only if components of VR based systems like head-mounted devices are compliant with standards, and are available to common masses. The road to widespread usage of VR systems is a challenge. Cost, technological capabilities, hardware-software co-design, learning curve, mass adaptation and usability analysis of VR systems are significant issues that need to be addressed. In this thesis, we focus on designing and implementing low-cost, customizable head-mounted devices for Virtual Reality applications. The cost aspect of the head-mounted device is addressed by making it application-specific (using degrees of freedom) and customizable (hardware and software development). As part of this work we studied currently available solutions and compared their sensing capabilities. The outcome resulted in choosing sensors based on the type of locomotion involved in the application and finalizing the hardware schematic. We further designed and fabricated the prototype to address VR applications with 3 degrees of freedom (roll, pitch and yaw). The application used to validate the prototype lies in Optometry domain. We propose a head-mounted device named CHORD (Customizable Head-mounted device for Occular disoRder Detection) to detect visual acuity disorders using VR applications. CHORD uses a standalone system with customizable VR scenes to detect myopic vision, astigmatism, colour vision deficiency and age-related macular degeneration. The product is validated using an empirical study with results showing an average accuracy of 85% in the tests conducted.

Abstract

Natural Language Processing (NLP) is a rapidly growing field focusing on the interaction between computers and human languages. It involves utilizing computational techniques to understand and generate natural language text. Several NLP tasks, such as question answering, text summarization, and machine translation, are widely researched for many languages, including Indian languages. Indian languages are resource-scarce and have distinctive characteristics posing different challenges for NLP. Recent advancements in NLP have helped in the development of models and techniques that are specific to Indian languages. However, for Telugu, a south Indian language, a lot of research is still needed to improve the performance of several NLP systems. Progress of such systems will benefit the huge Telugu-speaking community worldwide to communicate and access information in Telugu through various NLP applications. This motivates us to develop essential NLP resources and systems for Telugu. Firstly, the thesis provides an overview of the fundamental concepts and techniques used in NLP. Then we approach three specific NLP tasks: paraphrasing, question answering, and spelling correction. For these tasks, we address the problem of resource scarcity and then present the techniques to create the data for such low-resource languages. In this thesis, we have presented paraphrasing, question-answering, and spelling correction resources for the Telugu language. For paraphrasing, we presented two manually created and annotated corpora of size 1544 and 10000+ samples, respectively. We have also discussed the necessity for manual intervention while creating such resources. We have introduced a Telugu Question Answering Dataset - TeQuAD, with a size of 82k parallel triples. We also proposed the guidelines and methodologies that can be followed to create a Question Answering dataset for low-resource languages. We presented a Spell correction system for the Telugu language with the help of a synthetically created dataset.

Abstract

The applications of shared autonomy or human-robot interaction are growing rapidly in the field of autonomous robotics. Assisting human beings in dynamically changing environments in urban areas is still an active area of research. In crowded scenarios, in a structured environment such as public places with occlusions and dynamic obstacles, moving vehicles, people and so on - is the most critical part of this challenge. And, our work focuses on developing effective control strategies using model predictive control (MPC) because it is best known for handling such uncertainty and complex system dynamics relatively easily. While the extensive use of data-driven techniques using machine learning has become the de facto solution today, the underlying physics, the model of a system, and its behaviour are neces- sary to develop control laws. We first design an innovative MPC controller for a social person follower that can move safely around humans. We further incorporate motion-planning, target-tracking, and so- cial norms into a single holistic framework, being the first of its kind on a differential drive-wheeled mobile robot. To develop this robust person following behavior, we also employ path prediction using LSTM (Long-Short Term Memory) a type of recurrent neural network for supervised learning. This allows us for out-of-sight tracking and natural anticipation of a person’s future state. We also develop a local-map-based early relocation (ER) strategy that can reduce oscillations in the path, maintain the field of view (FOV) for long-term indoor navigation. Thus, we move beyond trivial person following to anticipating future visit locations and following them in the present. Overall, a non-linear MPC-based control law is designed using an online optimization problem with constraints on both kinematics and dynamics, as well as social norms of safety around humans. We implement these using 2D simulations in Matlab, and in Python to test the controller performance, runtime analysis, and error analysis. We show that the MPC framework can run in real-time with an adequate margin for adapting to changing human movement patterns, and agile enough for its changing movement speeds.

Abstract

In the research field of Artificial Neural Networks, there are ongoing efforts to address the issue of noise corruption in patterns. Various approaches and techniques have been explored to suppress noise and improve the accuracy of pattern classification. One of the research themes in this direction is the theory of Support Vector Machines [52] (SVM). SVMs aim to find an optimal hyperplane that maximizes the margin between classes, effectively reducing the impact of noise on classification. This approach formulates the problem of noise suppression as a quadratic programming problem. Hopfield Associative Memory [5] is another research area that has been investigated for noise suppression. The concept of null vectors of perceptron’s, which refers to vectors in the null space of the perceptron's transformation matrix, is introduced and utilized to suppress noise. By studying the null vectors of Extreme Machine Learning [49], researchers have explored their potential for noise suppression in pattern classification tasks. In addition to these approaches, the idea of stacking associative memories has been utilized to develop Hopfield neural networks [5] capable of suppressing noise in different dimensions. This includes 1-D, 2-D, and 3-D Hopfield neural networks that are designed to effectively handle noise corruption in patterns of various dimensions. Furthermore, to further enhance classification accuracy by suppressing noise, a real-valued Hopfield neural network based on a novel model of artificial neuron called the ceiling neuron has been proposed and studied. This innovation aims to improve noise suppression capabilities and overall performance in pattern classification tasks. Overall, this thesis explores and innovates different approaches to suppress noise in pattern classification using various neural network models, such as Support Vector Machines, Hopfield Associative Memory, and real-valued Hopfield neural networks based on ceiling neurons. These efforts contribute to the ongoing research in noise suppression techniques and aim to improve the accuracy and reliability of pattern classification systems.

Abstract

Distributed data analytics engines are designed to process and analyze large data sets in a distributed environment. The architecture of these engines is based on two key components: a distributed file system and a distributed computing framework. In this thesis, we consider the following two problems, coded data rebalancing in distributed file systems and coded distributed computing. For the data rebalancing problem, we consider replication-based distributed storage systems in which each node stores the same quantum of data and each data bit stored has the same replication factor across the nodes. Such systems are referred to as balanced distributed databases. When existing nodes leave or new nodes are added to this system, the balanced nature of the database is lost, either due to the reduction in the replication factor, or the non-uniformity of the storage at the nodes. This triggers a rebalancing algorithm, that exchanges data between the nodes so that the balance of the database is reinstated. The goal is then to design rebalancing schemes with minimal communication load. In a recent work [1] by Krishnan et al., coded transmissions were used to rebalance a carefully designed distributed database from a node removal or addition. These coded rebalancing schemes have optimal communication load, however, require the file-size to be at least exponential in the system parameters. In this work, we consider a cyclic balanced database (where data is cyclically placed in the system nodes) and present coded rebalancing schemes for node removal and addition in such a database. These databases (and the associated rebalancing schemes) require the file-size to be only cubic in the number of nodes in the system. We bound the advantage of our node removal rebalancing scheme over the uncoded scheme, and show that our scheme has a smaller communication load. In the node addition scenario, the rebalancing scheme presented is a simple uncoded scheme, which we show has optimal load. For the distributed computing problem, we consider the widely popular MapReduce distributed computing framework, which consists of three phases, namely Map, Shuffle, and Reduce. In previous works by Li et al. [2, 3], an optimal coded distributed computing scheme has been presented. This optimal scheme however requires very high file complexity, i.e., exponential in the number of servers K. To address this issue, low complexity distributed computing schemes using binary matrices derived from combinatorial designs have been presented in [4, 5]. In our work, we use similar principles to construct a distributed computing scheme via subspace designs, the q-analogs of combinatorial designs. While the scheme requires low file complexity, it has a higher communication load, when compared to the optimal scheme with equivalent parameters. Also, it is primarily useful for large local storage scenarios. Moreover, we also provide numerical comparisons with some existing baseline schemes.

Abstract

In the last few years, implantable medical devices are being used for the treatment of a growing number of pathologies. Implantable medical devices are sophisticated electronic circuits that are implanted into the body to restore or improve various body functions in patients. These devices can perform a range of tasks including sensing, control, and stimulation, which enables them to monitor and regulate the health of the patient. The electronics of a general biomedical device typically consist of several subsystems that work together to perform the desired functions. For example, energy delivery subsys-tem, analog-to-digital conversion subsystem, signal processing subsystem, communication subsystem. The stringent energy constraints dominate the design of biomedical systems. Additionally, systems powered through energy harvesting must be able to work with lower supply voltages. Low power consumption, low supply voltage and accuracy are few of the most important parameters for bio-implants. With all these constraints in play, we have put efforts to design current reference and voltage reference as they are present in almost every analog and mixed signal system. Efforts were put to design a low power current reference with high line sensitivity using peaking current sourceby exploiting the back-gate effect of a MOSFET in subthreshold region. The design works from a 0.55V supply voltage and generates a reference current of 5.6nA. A low line sensitivity of 0.022%/V is achieved and the design works for a temperature range of 0-80◦C. This current reference showsthe best result of line sensitivity among the previous state-of-the-art works. Consequently, we have come up with a switched capacitor network-based bandgap voltage reference (BGR) circuit designed to achieve high accuracy and low power consumption for implantable biomedical applications. A low- power clock generator circuit is proposed, to reduce the leakage current thereby reducing the circuit’s power consumption to 18.5nW at typical conditions. The design works from a supply voltage of 0.5V and has a TC of 74.5ppm/◦C over a temperature range of 0-80◦C. Lastly, we have expanded our research to the field of RF circuits for biomedical applications. An on-chip Vector Network Analyzer (VNA) isdesigned to detect a biomolecule for medical diagnosis. This technique, known as RF sensing, can be used to detect changes in the dielectric properties of a sample, which can indicate the presence of certainbiomolecule. In this part of the research, various blocks of the onchip VNA is demonstrated and the future scope of this work is explained.

Abstract

When volcanoes erupt explosively, the ash gets airborne and reaches till stratosphere vertically, and then spreads laterally to synoptic scales due to wind. Detecting the presence of ash in the atmosphere accurately is a challenge. Although several remote, in-situ, and near-sensing techniques exist, due to the variety and complexity of the physical and chemical properties of ash, that get ejected, even within between spells of the eruption of a given volcanic event, it is hard to distinguish it from other aerosols such as desert sand, ice clouds, etc. The false positives in the detection render these solutions unreliable and inconsistent. As a result, weather parameters are explored as an alternative strategy to predict the presence of ash. In specific, the temperature variable is identified as the proxy variable to study the spatial distribution of ash in the atmosphere. There are several beneficial reasons to study temperature because the values do not vary randomly, low-cost equipment suffice to gather the data, the diurnal variations can be accounted for easily, the ability to convert scales from negative to positive metrics for numerical calculations does not vary significantly with ash type, etc. For this research, the eruption of the Icelandic volcano, Eyjafjallajokull is chosen, due to the severe negative impact it created on the economy across Europe in April and May 2010. World Meteorological Organisation (WMO) and International Civil Aviation Authority (ICAO) together have created Volcanic Ash Advisory Center (VAAC) to model the concentration and simulate the transportation of ash to inform the hazards of ash fall to various stakeholders across the world. The London VAAC used a VAFTDM known as Numerical Atmosphericdispersion Modeling Environment (NAME) to model the ash spread. This theoretical model had several limitations, chief of them being related to the accuracy of the Numerical Weather Prediction (NWP) models that were supported for ash modeling. The UK Met Office dealing with the NWP associated with the NAME model supported and offered multiple models such as Unified Model (UM), ECMWF, and a few others to predict the weather variables. Since each VAAC uses its own NWP model that varies spatially and temporally, a benchmarking exercise was conducted to compare the Volcanic Ash Forecast Transport and Dispersion (VAFTD) models. The benchmarking allowed the use of either NCEP or ECMWF NWP. For this research, NCEP NWP was chosen for analysis since 6 out of 12 VAFTD models used this NWP.

Abstract

Microfluidic systems are effective tools for carrying out a wide range of chemical and biological analyses. Utilizing colorimetry in conjunction with microfluidic devices has proven to be a highly effective method for conducting various analyses that require rapid results. Such systems have significant applications in lab-on-a-chip systems and other medical devices that necessitate a colorimetric measurement to obtain quantitative results about the chemical composition of a given sample. Currently, most state-of-the-art techniques for performing microfluidic color detection use expensive cameras to study the captured image or are paper-based analytical devices that use capillary action to analyze color information. However, paper-based microfluidic devices are limited by the evaporation of test samples during experiments and therefore require high volumes of samples. Additionally, the surface tension of the liquid sample plays a crucial role in this process, and if not properly managed, the device may not provide accurate results. To overcome these limitations, a low-cost automated color detection system that could be used in conjunction with planar microfluidic pumps is presented in this work. The color detection system uses a light-emitting diode (LED) and a light-dependent resistor (LDR) to obtain color information. The LED is used to sequentially shine the primary color component wavelengths - red, green, blue - on the sample, and the color information is obtained from the light intensity incident onto the LDR after reflecting from the sample. The results showed that the system is accurate up to 92% and could be readily readily used to conduct different types of chemical analyses using colorimetry and can be employed in various industries to automate chemical and other biomedical processes completely. The flexible, planar micropumps discussed in this work are based on a nozzle-diffuser design that employs a central chamber connected to two trapezoidal flow diodes. The pump was fabricated using polydimethylsiloxane (PDMS) as a two-layer structure: a bottom layer with molded channels and the chamber, encapsulated by a planar top layer. The transparency of the pump and biocompatibility of PDMS make it an excellent material for biomedical and colorimetric applications. The pump design was characterized for its performance, both when placed on a flat surface and surfaces of different bending radii, by counting the number of compression cycles required for a known volume of fluid, here water, to flow from the inlet to the outlet. The pump produced a flow rate per compression cycle of 5.53 µl on a flat surface and that of 4µl, 3.6 µl when placed on surfaces of bending radii of 71cm and 45 cm, respectively. Finally, the pumping mechanism was completely automated by replacing the diaphragm with a dielectric elastomer actuator (DEA) and using a tesla valve for fluidic channels. The pressure difference imparted by the deformation of the actuator when a sinusoidal voltage is applied is converted to fluid flow (velocity). We also propose a microfluidic system that employs three microfluidic pumps and a central color detection module to perform effective microfluidic color detection. The proposed system provides a cost-effective solution for conducting chemical analyses using colorimetric assays.

Abstract

The estimation of the state of dynamic systems using measurements is a frequently arising challenge. State estimation is used in applications like robotics, industrial manufacturing, weather forecasting, target tracking, etc. The hidden state in the dynamic systems can be estimated using the recursive Bayesian filters. These filters estimate the state by tracking the posterior distribution. These filters include two steps, prediction based on solving the equation modeling the state evolution and update of the state using the measurements. Most real-world systems have non-linear dynamic and measurement models. Additionally, both models can be affected by arbitrary noise sources, which have either Gaussian or nonGaussian characteristics. For state estimation in linear Gaussian models, Kalman filters are employed. Extended Kalman filters (EKF) are utilized for state estimation in non-linear Gaussian models. For state estimation in non-linear non-Gaussian models, particle filters are generally employed. Particle filters, however, are ineffective in high-dimensional non-linear non-Gaussian models because they experience weight degeneracy. In this thesis, we propose incorporating the invertible particle flow methods in the Gaussian particle filter (GPF) framework to generate a proposal distribution. These methods are derived under Gaussian assumptions for the flow equation. The resulting particle flow Gaussian particle filter (PFGPF) method preserves the asymptotic characteristics of Gaussian particle filters and has the potential to perform better at state estimation in high-dimensional spaces. Secondly, we construct a particle flow Gaussian sum particle filter (PFGSPF), which roughly approximates the predictive and posterior as Gaussian mixture models, using a bank of PFGPF filters. In complicated estimating issues where a simple Gaussian approximation is insufficient, this approximation is helpful. We compare the performance of the proposed filters with the existing particle flow filters and particle flow particle filters (PFPF) in high dimensional complex numerical simulations.