Abstract

Field Programmable Gate Array (FPGA) is gaining popularity in robotics applications due to low cost, portability, low power dissipation, high speed and seamless interface to the hardware. On the other hand, robotics algorithms are mostly arithmetic in nature so they can be easily parallelized and implemented on FPGA. Asynchronous FPGAs have gained remarkable research interest over the last decade.FPGA exceeds the computing power of software-based implementations by breaking the paradigm of sequential execution and accomplishing more per clock cycle by enabling hardware level parallelization at an architectural level. Introducing parallel architectures for a computationally intensive algorithm like Rapidly-Exploring Random Trees(RRT) will result in an exploration that is fast, dense and uniform. A rapidly-exploring random tree (RRT) is an algorithm designed to efficiently search non-convex,high-dimensional spaces by randomly building a space-filling tree. The tree is constructed incrementally from samples drawn randomly from the search space and is inherently biased to grow towards large unsearched areas of the problem. But the overall performance of the algorithm comes down due to high computation delay with the increase in the maps size and geometric constraints. One way to improve the performance is by parallelizing RRT. Here parallelizing RRT refers to multiple RRT modules working in parallel to achieve faster exploration of the map. And to accommodate all the data that is coming from multiple modules an architecture is required. So with the aim to minimize computation delay with the increase in maps size and geometric con- straints, we present the FPGA based hybrid architecture that allows multiple RRTs to work together combined with modified K-means sensitive to the geometric and kinematic constraints of the robot for accelerated and uniform exploration of the map. Through a cost function delineated in later sections, FPGA based combinatorial architecture delivers superlative speed-up but consumes very high power while hierarchical architecture delivers relatively lower speed-up with acceptable power consumption levels. The hybrid architecture combines the qualities of both combinatorial and hierarchical architecture. To determine the number of RRT nodes to be allotted to the combinatorial and hierarchical blocks of the hybrid architecture, a cost function, comprised of fundamentally inversely related speed-up and power parameters, is formulated. This maximization of the cost function, with its associated constraints, is then mathematically solved using a modified branch and bound, that leads to the optimal allocation of RRT modules to both blocks. For the initial placement of RRTs, we came up with a software-defined strategy for identification of these starting points using a modified K-Means. Strategized placing of the seeds (starting points) of each RRT module can lead to a more homogeneous and faster exploration as compared to the indiscriminate placement of the same. Further, to validate the empirical working of the architectures in a real-world scenario, a Zynq-SoC based differential drive robotic system was built as a test platform to run the architectures in real time in a complex environment filled with obstacles. It is observed that this hybrid architecture delivers the highest performance-per-watt out of the three architectures for differential, quad-copter and fixed-wing kinematics. The overall system is a hardware-software co-design to achieve real-time, ASIC like performance that delivers accelerated and homogeneous exploration.

Abstract

Field Programmable Gate Array (FPGA) is gaining popularity in robotics applications due to low cost, portability, low power dissipation, high speed and seamless interface to the hardware. On the other hand, robotics algorithms are mostly arithmetic in nature so they can be easily parallelized and imple- mented on FPGA. Asynchronous FPGAs have gained remarkable research interest over the last decade. FPGA exceeds the computing power of software-based implementations by breaking the paradigm of sequential execution and accomplishing more per clock cycle by enabling hardware level parallelization at an architectural level. Introducing parallel architectures for a computationally intensive algorithm like Rapidly-Exploring Random Trees(RRT) will result in an exploration that is fast, dense and uniform.A rapidly-exploring random tree (RRT) is an algorithm designed to efficiently search non-convex, high-dimensional spaces by randomly building a space-filling tree. The tree is constructed incrementally from samples drawn randomly from the search space and is inherently biased to grow towards large unsearched areas of the problem. But the overall performance of the algorithm comes down due to high computation delay with the increase in the maps size and geometric constraints. One way to improve the performance is by parallelizing RRT. Here parallelizing RRT refers to multiple RRT modules working in parallel to achieve faster exploration of the map. And to accommodate all the data that is coming from multiple modules an architecture is required. So with the aim to minimize computation delay with the increase in maps size and geometric constraints, we present the FPGA based hybrid architecture that allows multiple RRTs to work together combined with modified K-means sensitive to the geometric and kinematic constraints of the robot for accelerated and uniform exploration of the map. Through a cost function delineated in later sections, FPGA based combinatorial architecture delivers superlative speed-up but consumes very high power while hierarchical architecture delivers relatively lower speed-up with acceptable power consumption levels. The hybrid architecture combines the qualities of both combinatorial and hierarchical architec-ture.To determine the number of RRT nodes to be allotted to the combinatorial and hierarchical blocksof the hybrid architecture, a cost function, comprised of fundamentally inversely related speed-up and power parameters, is formulated. This maximization of the cost function, with its associated constraints, is then mathematically solved using a modified branch and bound, that leads to the optimal allocation of RRT modules to both blocks. For the initial placement of RRTs, we came up with a software-defined strategy for identification of these starting points using a modified K-Means. Strategized placing of the seeds (starting points) of each RRT module can lead to a more homogeneous and faster exploration as compared to the indiscriminate placement of the same. Further, to validate the empirical working of the architectures in a real-world scenario, a Zynq-SoC based differential drive robotic system was built as a test platform to run the architectures in real time in a complex environment filled with obstacles. It is observed that this hybrid architecture delivers the highest performance-per-watt out of the three architectures for differential, quad-copter and fixed-wing kinematics. The overall system is a hardware-software co-design to achieve real-time, ASIC like performance that delivers accelerated and homogeneous exploration.

Abstract

There are different aspects of daylight that has been proven to effect the human visual and non-visual system. Daylighting in buildings have a substantial impact on workers’ health, productivity, performance, and circadian system. It can enhance the overall experience of built spaces and benefit the inhabitants in many ways. However, it is necessary to control the daylight admission inside buildings to avoid visual and thermal discomfort. Automated shading and integrated lighting control systems are being used in buildings, for controlling the daylight either in closed loop or open loop system. A typical commercial closed loop system usually has a series of indoor photosensors integrated with dimmers to maintain desk illuminance levels. Unfortunately, these sensors are expensive to install at each desk, challenging to commission and difficult to calibrate therefore highly prone to errors. Whereas in open loop systems, sensors are mounted at external surface of façade to measure sky conditions and daylight availability. However, these devices are only capable of giving single specific measurements without accounting size, directionality, temporal and spatial dynamics of clouds. This leads to failure of the control system to respond in a timely fashion by inaccurately estimating visual discomfort parameters. A reliable estimation of sky conditions and indoor daylight metrics is crucial to open loop controls system. Currently, the daylighting and lighting design practices followed by architects and designers predominantly focuses on the visual characteristics of light. These characteristics include provision of sufficient work plane illuminance, colour rendition and appearance of lighting. There is little or no consideration imparted towards the circadian impacts of lighting and daylighting devices in design process. The impact of lighting in regulating the circadian rhythm in human body has been strengthen with the recent discovery of intrinsically photosensitive retinal ganglion cells (ipRGCs). However, there is lack of accessible instruments to measure the circadian lighting and therefore it’s difficult to monitor daylit spaces and incorporate circadian aspects of lighting in design process. To understand the impacts of lighting, it is essential to devise methods that can quantify both visual and non-visual parameters of daylighting. In this work, a system has been developed to operate a physical setup to calculate glare, illuminance, correlated colour temperature and circadian lighting in an indoor space using high dynamic range (HDR) imaging and control the window roller shades based on glare levels in an open loop system. This thesis gives the details of the system and its performance in various test beds. Following are the objectives: 1. Determine, validate and control simulation based indoor glare and illuminance levels using real time image-based sky maps. 2. Determine and validate correlated colour temperature and circadian lighting in two different experimental setups at two different locations:Results indicate that HDR image-based sky maps provide a good estimation of diffuse sky and take the surroundings, cloud dynamics into consideration as well. In the operational range, the errors observed are within 12% limit. However, larger errors are observed when lux is higher than 2000lux. These errors are of less importance since daylight control will be operational during that time. The circadian lighting measured in test setups vary significantly throughout the day. Lower daylight levels led to lower melanopic lux levels. The correlated colour temperature of the light provided by the tubular daylight device was generally found to be higher than that of the electric lights in test room. The HDR photography technique makes measuring circadian light accessible to all lighting professionals and facilitates field measurements of circadian lighting.

Abstract

Autism Spectrum Disorder (ASD) is a clinical umbrella term for neuro-developmental disorders that are characterized by atypical social behavior, including deficits in receptive and expressive language, theory of mind (TOM), and cognitive flexibility deficit. It encompasses autism, Asperger’s syndrome, pervasive developmental disorder not otherwise specified (PDD-NOS), and childhood dis-integrative disorder. Accumulating evidence indicates that ASD is associated with alterations of neural circuitry compared to typically developing (TD) individuals. Despite numerous studies, there is a lack of consensus on the nature of atypical functional connectivity in ASD. We hypothesize that studying the effects of age, disease and their interaction together using resting-state functional MRI (rs-fMRI) data can reveal crucial insights about the autistic brain. fMRI measures blood oxygen level dependent (BOLD) signals that indirectly maps the neuronal activity across different regions of the brain. In this thesis, we studied these effects using two paradigms: Firstly, we studied the effects of age, disease and their interaction on static functional connectivity by treating the brain as a complex graph in children and adolescents (9-16 years). Graph-theoretic analyses were performed on the functional brain network. We characterized the regions of interest (ROIs), connections and functional networks that are altered in development and in ASD. The multiple network measures representing the modular organization (integration and segregation) in the whole-brain were quantified. The results showed that ASD exhibits increased functional integration at the expense of decreased functional segregation. In ASD adolescents, there is significant decrease in modularity suggesting a less robust modular organization and an increase in participation coefficient suggesting more random integration and widely distributed connection edges. There is significant hypo-connectivity observed in the adolescent group especially in the Default Mode Network (DMN) while the children group shows both hyper- and hypo-connectivity. This lends support to a model of global atypical connections and further identifies functional networks and regions that are independently affected by age, disease and their interaction. Secondly, we have also explored flexibility of the brain as a novel approach to characterize ASD using the temporal dynamics of BOLD signals. The dynamic variability in the connection strength and the modular organization in terms of the measures flexiblity, cohesion strength and disjointness were explored for each subject to characterize the effects of age, disease and their interaction. In ASD, we observed significantly higher inter-subject dynamic variability in connection strength as compared to TD. This hyper-variability relates to the symptom severity in ASD. We found that the whole-brain flex-ibility correlates with static modularity only in TD. Further, we observed a core-periphery organization in the resting-state, with Sensorimotor and Visual regions in the rigid core; and DMN and attention areas in the flexible periphery. TD also develops a more cohesive organization of sensorimotor areas. However, in ASD we found a strong positive correlation of symptom severity with flexibility of rigid areas and with disjointness of sensorimotor areas. The regions of the brain showing high predictive power of symptom severity were distributed across the cortex, with stronger bearings in the frontal, motor and occipital cortices. Our study demonstrates that the dynamic framework best characterizes the variability in ASD and overcomes the limitations of the analyses that consider ASD as a single group. Overall, this thesis adopts a systematic approach to study pathology and development using static as well as dynamic functional connectivity. These are very useful in characterizing the topology of the brain along with the patterns in transient connectivity.

Abstract

The rise in the use and penetration of Artificial Intelligence/ Machine Learning in everyday life has led to many complex issues in Law and Policy. Chiefly, the governance of AI entities under the ambit of law. This thesis explores two issues in detail: General Data Protection Regulation (GDPR) floated ‘right to explanation’ and the problematic nature of Intellectual Property for AI.With the advent of GDPR, the domain of explainable AI and model interpretability has gained addedimpetus. Methods to extract and communicate visibility into decision-making models have become a legal necessity. The GDPR terms this as the ‘Right to Explanation’: Any person affected by the decision of an autonomous decision making system is empowered to seek an answer for the same. This is especially relevant in 21st century scenarios wherein AI systems are used in determining the credit worthiness, evaluating stocks, recommending news, screen candidates for jobs etc. In all these scenarios, the end user suffers a social, economic or a psychological impact by the discretion of the machine. Most of these AI systems function with complex black box models which have very little visibility in its inner workings. Even its developers might not be fully aware as to why and how these self learning systems adapt and react to new data either due to their evolving nature or its sheer complexity. In such a scenario, it is apt that appropriate justifications for the actions of the machine be generated. This thesis explores two such methods of generating explanations: an Antehoc technique conceptualized atop the existing SUMO Ontology and a Posthoc method constructed out of model interpretability techniques which generate contrastive and counterfactual explanations. Our computational explanations are a remedy to the challenge posed by the ‘Right to explanation’ and also serve in making the blackbox models more fair, reliable and most importantly understandable for the end user. The next part of the thesis focuses on the problematic nature of Intellectual Property (Copyright & Patents) for AI. With the ‘third wave’ of Artificial Intelligence, there is a massive revival and upsurge in AI related product development. An important entity behind the AI architecture, the neural network needs to be studied carefully that adequate protection for its innovation can be secured. A key feature of the Neural Network, the Neural weights hold the inferential rules and knowledge, thus are a new way to embody knowledge and information, a new form of intellectual property to which IP laws will have to adapt. We present our discussion that sheds light on the nature of this innovation and brings to context why it is relevant to secure Intellectual Property for Neural Weights. We also rebase our arguments in the backdrop of the debates that were set off on this same topic in 1990. Our research traces the shape of this problem ever since its conception and brings to the fore the newer and expanded notions behindNeural Networks, AI and their place in the Copyright laws. We also propose tentative solutions and the hurdles in implementing them as part of the study on streamlining the rough contours of the subject. As with Copyright, Patent Laws for AI are also predated. The Patent law is silent about tackling cases when an AI develops something novel. In the context of the current technologies and AI engines, this problem is particularly ripe as AI driven innovation will cause a radical shift in the pace, quality and areas of innovation which will need a massive overhaul of existing laws which do not allow non-human innovation to be registered in the current innovation market. To address this problem, we introduce a policy framework to adjudge innovation such that both human and machine driven innovation co-exists and complements each others R&D efforts. Our framework ensures that the legal status of machine is unchanged, the innovation quality surges and frivolous patent applications are pruned in early. The intention of this thesis is to elaborate, explore and mitigate the rough edges between AI and Law. The former part of the thesis proposes computational measures towards increased accountability, fairness and transparency in systems and the latter is directed towards a more policy angle between AI and Intellectual Property. The summum bonum of AI is to integrate with the human fabric and catalyze human efforts. This is possible if trust and acceptance is legally and socially established in these systems. This work is a small step towards the giant leap of making AI safe, governable and usable for the benefit of humanity.

Abstract

India’s cooling demands are growing exponentially with rapid urbanization, increase in global warming due to change in climate and an increase in affordability and aspiration levels of rising number of people. Due to the availability of limited energy resources, meeting these increasing cooling demands is a big challenge. A mixed-mode building can operate both as a naturally ventilated building and air-conditioned (AC) building. It allows natural ventilation (NV) when the outside weather conditions are favorable and deploys air-conditioning when natural ventilation is not able to provide sufficient comfort. A well-designed mixed-mode building can prove to be one of the solutions to address this challenge. Therefore, there is a need to evaluate the energy saving potential if the building is operated in mixed-mode. Also, the current practice is to have static setpoint throughout the summers which is not energy-efficient. Based on the recent studies on adaptive thermal comfort, more energy savings can be achieved further if this setpoint can change according to outdoor conditions. In this thesis, a simulation-based study has been done to evaluate the potential of operating buildings in mixed-mode for five Indian climatic zones according to Energy Conservation Building Code (ECBC) 2017 (India) – hot-dry, warm-humid, composite, temperate and cold. Five representative cities were selected for each climatic zone namely Ahmedabad, Kolkata, Lucknow, Bangalore and Srinagar respectively. Initially, a methodology for climate potential of natural ventilation has been applied and results have been discussed. Two adaptive thermal comfort models of ASHRAE:55 Adaptive Model of Thermal Comfort (ASHRAE ATC) and Indian Model of Adaptive Comfort (IMAC) have been used to calculate the favourability of natural ventilation for the five climatic zones of India. Bangalore in temperate climate was observed to be the most suitable with almost 60% of hours favouring natural ventilation from a climatic point of view using IMAC. On the other hand, Kolkata was observed to be the least suitable with only 35% of hours in favour of natural ventilation. To estimate the cooling energy savings, a model for the mixed-mode building has been proposed and simulated in EnergyPlus for all the five locations for two indoor operative thermostat (OT) setpoint controls – fixed thermostat of 26℃ and adaptive thermostat based on IMAC. While evaluating the energy consumption of mixed-mode building, care has been taken to evaluate the sensitivity of NV to wind velocity. This is because weather files are developed by taking the values available at the airport and those wind velocities might not be the same in densely populated parts of the cities. Results have shown that significant cooling energy saving potential is present in both temperate and cold climate. Potential energy saving of 60% is obtainable for Srinagar (cold), 58% in Bangalore (temperate), 29% in Lucknow (composite), 26% in Ahmedabad (hot-dry) and 17% in Kolkata (Warm-humid) when deploying adaptive OT thermostat control according to IMAC Model. Since adaptive model offers higher flexibility range, potential for energy savings using an adaptive model is greater than a fixed thermostat model. A methodology is also proposed to enable real-time control of operable windows for mixed mode ventilation building. The methodology consists of using real-time weather forecast and predicting the internal loads to estimate the working of windows and correspondingly the mechanical cooling system, for the next timestep. The model is simulated in natural ventilation mode for the next timestep and checked if there is potential to save cooling energy due to inside conditions. Windows remain open if the inside conditions remain favourable, otherwise the mechanical cooling system is scheduled to ON and this repeats for the next window. The methodology has been shown using AERIS Weather Data for hourly forecast of weather parameters and deploying Seasonal Autoregressive Integrated Moving Average (SARIMA) Models to model a day-ahead prediction for the lighting load, electric load and the occupancy profiles. Co-simulation of EnergyPlus is carried out with Building Controls Virtual Test Bed (BCVTB) to enable real-time simulation wherein; the EnergyPlus weather file parameters and input building loads are updated every hour. Based on the indoor and outdoor conditions, EnergyPlus calculates the schedule for opening/closing of windows. A simulation-based case study for Hyderabad, India is performed to demonstrate the working of the proposed methodology. Reduction of 25% in cooling energy demand was estimated in a mixed mode building as compared to the fully air-conditioned building during the study period.

Abstract

Installation, deployment and maintenance of an NLP system, can be a daunting dask; based on the number and complexity of the components involved. One such system can be a hybrid Machine Translation system, composed of several modules which define the transformation of a given word, phrase or sentence, from one language to another. The end users of such a system can be developers themselves who want to improve it. To achieve this, the system as whole, needs to adopt an architecture which lets the users control and change the order of components in the pipeline, be able to intervene the execution in middle to debug one or more components, tweak inputs/outputs without having to rewrite the components, be able to easily replicate the system on their local box without having to worry about the hassle of compiling everything from source, be able to quickly replace a component with a higher or lower version and be able to see the impact on the final result quickly; basically make development iterative and fast. The system should also expose an interface for those users who want to build something on top of it, without having to worry about the internal details. The ideas proposed in this thesis try to cater to the needs of a broad category of users, in an attempt to keep their work-flow as simple as possible. We propose an architecture where we show how to identify and transform a monolithic system into small, individual components (each being a linguistic unit), identify bottlenecks from an operating system’s point of view, identify scalable components and finally provide an easy mechanism to interact with the system. To achieve this, we apply the proposed architecture over an existing system (Sampark MT System) and walk through it’s transformation. Towards the end, we show the creation of a web client which shows how easy it becomes to interact with the modified system. We also apply our proposition to show that it can be applied to any pipeline based system by thinking of it as a disconnected, directed acyclic graph. We also show how the modified system can be deployed on the cloud easily and how individual components can be scaled up or down as per needs. To be able to plan the overall architecture and produce guidelines for enabling large scale collaborative development, a structured systems analysis and design, from the point of view of both, a computational linguist and a systems engineer is required. This thesis provides the foundation in that direction by enhancing an existing system, reducing the overall runtime of it’s components by greater than 85%, improving the test-dev-deploy cycle for computational linguists and discussing a generalized architecture on top of which, further complex systems can be built for specific purposes.

Abstract

Systems biology focuses on the integration of experimental, mathematical and computational techniques to develop systemic views and predictive models of biological systems. Transcriptomics has enabled to obtain genome-wide measurements in health and disease. The analysis of transcriptomic data in its entirety allows detection of broad coordinated trends which cannot be discerned by more targeted assays. Biological network reconstructions are important tools in systems biology in order to model the behavior as a whole biological system. Biological networks can also be used as a scaffold for integrative analysis where high throughput data from different conditions can be integrated into the biological network to obtain condition-specific molecular biomarkers and mechanisms. In this thesis, we aimed to gain a systems-level understanding of two obesity-linked metabolic disorders: Non-alcoholic fatty liver disease (NAFLD) and Type 2 Diabetes (T2D) using transcriptomics data obtained from healthy and disease groups. NAFLD is a complex spectrum of diseases ranging from simple steatosis to Non-Alcoholic Steato hepatitis (NASH) with fibrosis, which can progress to cirrhosis and hepatocellular carcinoma. The pathogenesis of NAFLD is complex, involving crosstalk between multiple organs, cell-types, and environmental and genetic factors. Dysfunction of the adipose tissue plays a central role in NAFLD progression. Here, we analysed transcriptomics data obtained from the Visceral Adipose Tissue (VAT) of NAFLD patients to understand how the VAT metabolism is altered at the genome scale and co regulated with other cellular processes during the progression from obesity to NASH with fibrosis. For this purpose, we constructed weighted gene co-expression network and studied the organization of the disease transcriptome into functional modules of cellular processes and pathways. Our analysis revealed the coordination of metabolic and inflammatory modules (termed ”immunometabolism”) in the VAT of NAFLD patients. We found that genes of arachidonic acid, sphingolipid and glycosphingolipid metabolism were upregulated and co-expressed with genes of proinflammatory signalling pathways and hypoxia in NASH/NASH with fibrosis. We hypothesize that these metabolic alterations might play a role in sustaining VAT inflammation. Furthermore, immunometabolism related genes were also coexpressed with genes involved in Extracellular Matrix (ECM) degradation. Our analysis indicates that upregulation of both ECM degrading enzymes and their inhibitors (incoherent feedforward loop) poten- tially leads to the ECM deposition in the VAT of NASH with fibrosis patients.Further, we studied the pancreatic islets in health and T2D. Studies on how pancreatic islets respond under physiological and pathological conditions are obtained mostly based on the analysis of whole islet transcriptome. However, the measurement from the whole islets quantifies the average behaviour of dominant cell types, thereby making it difficult to understand the cell-type-specific changes. Recently, the advent of single-cell RNA sequencing (scRNA-seq) technique has generated valuable resource on islet biology and T2D. This provides an opportunity to understand the different cell types/states at both the network and individual gene expression levels. Here, we inferred the gene regulatory networks (GRNs) of pancreatic cells from scRNA-seq data in healthy and T2D. Clustering of cells based on GRNs identifies endocrine and exocrine cells and multiple stable cell states in each alpha, beta and ductal cells. The phenotypic variations in cell states due to obesity and T2D are indistinguishable. Therefore, the trajectory of cells in pseudotime was constructed based on the cell-type-specific gene expression. The analysis shows that continuous spectrum of cell states exists with phenotypic-dependent branching and donor cellcell variability in endocrine and exocrine cell types. We characterized the genes that give rise to bifurcation in the trajectory. Our study demonstrates that the network and trajectory inference approaches can be used to better understand the behaviour of pancreatic cells in health and disease. Overall, this thesis generates tissue- and cell-type specific networks that can be further used to understand the underlying pathology of metabolic disorders.

Abstract

Social media is becoming increasingly popular as a platform for sharing personal health-related information. This information can be utilized for public health monitoring tasks, particularly for phar- macovigilance, via the use of natural language processing (NLP) techniques. However, the language in social media is highly informal, and user-expressed medical concepts are often nontechnical, descriptive, and challenging to extract. This poses a significant challenge to the existing Natural Language Processing (NLP) and Information Retrieval (IR) methods. In this thesis, we try to address the problem of “Adverse Drug Reaction (ADR) Mention Extraction”, a specific form of pharamacovigilance, from social media. Specifically, we try to address the problem in the setting where we have a limited access to labeled data but potentially unlimited access to unlabeled data, which is often the case in the real world. We investigate the applicability of two of the popular and well-studied approaches for handling such problems: Semi-supervised learning and Multi-task learning.In the first approach, we propose a novel semi-supervised method which utilizes a novel task of “Unsupervised Drug-Name Prediction”, which is the prediction of drug-name from the tweet context where the actual drug-name is identified and masked from the tweet and the objective is to predict the drug-name from the masked tweet. Enforcing the network to predict the drug-name allows it to learn the context around the drug-name mention. Through experiments we demonstrate that pre-training the network with such task helps in improving performance on the main task. In the second work, we make use of more traditional and well-studied semi-supervised method: co-training, for ADR extraction. To the best of our knowledge, we are the first ones to apply a co-training based methods for entity extraction using deep learning. Through experiments we demonstrate that co-training based network gives a significant improvement in the main task as compared to its fully supervised counterpart. In the third work, we propose a multi-task learning based method which can utilize a similar auxiliary task (adverse drug event detection) to enhance the performance of the main task of our interest, i.e., ADR extraction. Further, in the absence of auxiliary task dataset, we propose a novel joint multi-task learning method to automatically generate weak supervision dataset for the auxiliary task when a large pool of unlabeled tweets is available. Experiments with a large unlabeled pool of tweets, we show that the proposed approach outperforms the state-of-the-art methods for the ADR mention extraction task by ∼7.2% in terms of F1 score.

Abstract

Alzheimer's disease (AD) is a neurodegenerative disorder contributing to a rapid decline in cognitive function and ultimately dementia. Most cases of AD occur in elderly and later years. There is a growing need for understanding the relationship between aging and AD to identify shared and unique hallmarks associated with the disease in a region and cell-type specific manner. Although genomic studies on AD have been performed extensively, the molecular mechanism of disease progression is still not clear. The major objective of our study is to obtain a higher-order network-level understanding of aging and AD, and their relationship using gene expression profiles of young (20–50 years), aging (70–99 years), and AD (70–99 years). Also, these approaches were used to determine regional vulnerability in aging and AD. Hippocampus (HC) is vulnerable to damage at early stages of AD and altered neurogenesis in the hippocampus is linked to the onset of AD. We combined the weighted gene co-expression network and weighted protein-protein interaction network-level approaches to study the transition from young to aging to AD in HC. We found that modules associated with astrocytes, endothelial cells and microglial cells are upregulated and significantly correlate with both aging and AD. The modules associated with neurons, mitochondria and endoplasmic reticulum are downregulated and significantly correlate with AD than aging. The oligodendrocytes module does not show significant correlation with either aging or disease. Further, we identified aging- and AD-specific interactions/subnetworks by integrating the gene expression with a human protein-protein interaction network. We found dysregulation of genes encoding protein kinases (FYN, SYK, SRC, PKC, MAPK1, ephrin receptors) and transcription factors (FOS, STAT3, CEBPB, MYC, NFKβ, and EGR1) in AD. Further, we found genes that encode proteins with neuroprotective function (14- 3-3 proteins, PIN1, ATXN1, BDNF, VEGFA) to be part of the downregulated AD subnetwork.Further, network analysis on regions Entorhinal Cortex (EC), Post Central Gyrus (PCG) and Superior Frontal Gyrus (SFG) revealed similar cell-types were affected in aging and AD but the extent of changes is more diverse in aging than AD. Glial and endothelial changes were prominent in aging whereas neuronal changes were prominent in AD. Synaptic transmission, protein processing, mitochondria energy metabolism, and inflammation were the major processes affected in these regions. In aging, EC showed minimal changes followed by HC and PCG while SFG showed maximum changes. In AD, EC and HC showed maximum changes while PCG and SFG showed minimal changes. Overall, the network-level analysis showed differences in regional vulnerability to aging and AD. Taken together, our study highlights that simultaneously analyzing aging and AD will help to understand the pre-clinical and clinical phase of AD and aid in developing the treatment strategies.