Abstract

Remote labs allow students from anywhere in the world to access and conduct experiments without the need to physically be present in a lab at anytime. This is extremely important for students with little to no access to proper science labs because of a lack of infrastructure or a pandemic. A major open problem statement is adding scalability to existing Remote Labs, for several reasons, including the ability to handle growing demand while reducing costs and increasing flexibility. There is a potential for scaling up the process so that queue delay can be removed and people are able to access dedicated experiment setups. The work presented in this thesis is aimed at scalability of Remote Labs by using miniaturization and partial streams. Miniaturization involves making the experiment apparatus simple, smaller and portable. Partial streams involves using a single camera for creating point-of-view video streams for multiple units of the same experiment apparatus. Both aspects operate in tandem in order to be able to scale up the in-house Remote Labs system. The proof of concept demonstration of the methodology put forward is conducted using two miniaturized setups of the school-level experiment “Vanishing Rod”. These experiments are set up in our Remote Labs at IIIT Hyderabad, India (IIIT-H). The miniaturized setups are 3D printed in the lab. To demonstrate the effectiveness of the proposed approach, a performance comparison in terms of cost, size, and energy consumption is carried out for the proposed architecture (miniaturized + partial streaming) compared to the traditional setup (lab-scale + single-streaming). In both cases, the software framework for the dashboard is developed and implemented at IIIT-H. With the above approach, power cost is reduced to one-sixth and actuation component cost is reduced to one-fifth. In addition, the miniature setups require less space as compared to the lab-scale setups, making it easier to install.

Abstract

Understanding the semantics of the scene to automate the decision process for self-driving cars completely is becoming a crucial task to solve in computer vision. Due to the recent progress in the state of autonomous driving, added with a lot of semantic segmentation datasets for road scene understanding being proposed, semantic segmentation of road scenes has recently evolved to be an important problem to tackle. But training semantic segmentation models becomes a resource-intensive task since it requires multi-GPU training and therefore becomes the bottleneck to reproducing results for better understanding quickly. This thesis introduces challenges and provides solutions to reduce the training time of segmentation models by introducing two small-scale datasets. Additionally, the thesis explores the potential of employing neural architecture search and automatic pruning techniques to create efficient segmentation modules in resource-constrained settings. Chapter2 of the thesis introduces the problem of semantic segmentation and discusses some deep learning approaches to solve supervised semantic segmentation. We briefly discuss the different metrics used and also touch upon the statistics of various datasets that are available in the literature to train semantic segmentation models. Chapter 3 of the thesis explains the need of having a dataset based on the Indian road scenario. Most of the datasets in the literature are captured in Western settings having well-defined traffic participants, delineated boundaries, etc, which seldom mold in the Indian setting. We describe the annotation pipeline, along with the quality check framework used to annotate the dataset. Now, though the IDD dataset [121] caters to the Indian setting, this dataset is still quite resource intensive in terms of GPU computation. Hence, there is a need to have a small resolution, less label-sized dataset for rapid prototyping. We introduce our proposed datasets and provide a detailed set of experiments, and statistical comparisons with the existing datasets to substantiate our claim regarding the usefulness of the proposed solution. We also show through experiments that the models trained using our datasets can be deployed on low-resource hardware such as Raspberry Pi. At the end of this chapter, we also look into the significance of the proposed datasets in facilitating challenges at two prominent conferences: the International Conference on Computer Vision (ICCV) and the National Conference on Pattern Recognition, Image Processing, and Graphics (NCVPRIPG) in 2019. These challenges aimed to address semantic segmentation in resource-constrained settings, inviting innovative architectures capable of achieving decent accuracy on these proposed datasets. We also discuss the potential application of these datasets in teaching semantic segmentation through a course of notebooks introducing traditional as well as deep learning-based methods to perform segmentation. These notebooks are plug-and-play, where the first three notebooks can run on laptop CPU, while the fourth notebook requires GPU access.

Abstract

The conventional method of manually reading analog meters to track consumption trends is both laborious and costly. Moreover, it falls short in effectively managing sustainable water supplies, neces- sitating accurate monitoring techniques to provide real-time insights into water usage for consumers. While digital water meters have been introduced, their high cost makes them impractical for widespread adoption, and they lack analytical capabilities for interpreting consumption patterns. In contrast, tradi- tional analog water meters boast simplicity, low power consumption, durability, and reliability, but they still rely on manual readings, which is inconvenient. To address this issue, an automated data-capturing system is required to transmit real-time meter readings to a cloud server. Smart water meters, powered by robust machine learning (ML) and deep learning (DL) algorithms for meter reading detection, can analyze the data collected to gain valuable insights into water consumption patterns and detect leaks, facilitating more efficient water management. Ultimately, smart water monitoring devices can empower users to reduce their water usage and contribute to water conservation efforts. Keeping all these aspects in mind, the thesis can be divided into three parts: Firstly, this thesis introduces an IoT based economic retrofitting setup for digitising the analog water meters to make them smart. The setup contains a Raspberry-Pi microcontroller and a Pi-camera mounted on top of the analog water meter to take its images. The captured images are then preprocessed to estimate readings using a ML model. The employed ML algorithm is trained on a rich dataset that includes digits from the images of water meters captured by the hardware setup for ten days. The readings are posted on a cloud server in real-time using Raspberry-Pi. High temporal resolution plots of flow rate and volume are generated to derive inferences. The collected data can be used for deriving water consumption patterns and fault detection for efficient water management. After that, the thesis proposes a DL-based algorithm which is used for improving the performance of digit detection from IoT-based analog water meters. The DL algorithm is trained on a rich dataset of over 160,000 images collected from six water nodes deployed at locations with different environmental conditions. A detailed comparison between the proposed DL and ML algorithm is made based on de- tection accuracy, feature analysis, error analysis, and computational complexity analysis. It is observed that compared to the ML model, the proposed DL model maintained a higher detection accuracy and is more generalized in terms of feature extraction, which makes the algorithm robust. Finally, the thesis presents a comprehensive analysis of water supply behaviour on an educational campus, focusing on two distinct regions: student hostels and faculty/staff quarters. The investigation delves into the impact of water supply patterns on a monthly and weekly basis. Notably, it highlights how each month, with its unique characteristics such as holidays, exams, and class schedules, influences the water supply in both regions. One key difference between the two regions is that students reside in one, leading to significant variations in water usage based on the number of holidays. Conversely, the other region accommodates families, resulting in a consistent water requirement regardless of col- lege holidays. The findings from this analysis are crucial for understanding water distribution patterns, particularly within intermittent water supply (IWS) systems, with the ultimate goal of enhancing the efficiency and robustness of water distribution. By thoroughly examining the water supply behaviour in an educational campus and considering various factors that influence it, this work contributes to a better understanding of water management on campus.

Abstract

Imparting machines the capability to understand documents like humans do is an AI-complete problem since it involves multiple sub-tasks such as reading unstructured and structured text, understanding graphics and natural images, interpreting visual elements such as tables and plots, and parsing the layout and logical structure of the whole document. Except for a small percentage of documents in structured electronic formats, a majority of the documents used today, such as documents in physical mediums, born-digital documents in image formats, and electronic documents like PDFs, are not readily machine readable. A paper-based document can easily be converted into a bitmap image using a flatbed scanner or a digital camera. Consequently, machine understanding of documents in practice requires algorithms and systems that can process document images—a digital image of a document. Successful application of deep learning-based methods and use of large-scale datasets significantly improved the performance of various sub-tasks that constitute the larger problem of machine understanding of document images. Deep-learning based techniques have successfully been applied to the detection, and recognition of text and detection and recognition of various document sub-structures such as forms and tables. However, owing to the diversity of documents in terms of language, modality of text present (typewritten, printed, handwritten or born-digital), images and graphics (photographs, computer graphics, tables, visualizations, and pictograms), layout and other visual cues, building generic-solutions to the problem of machine understanding of document images is a challenging task. In this thesis, we address some of the challenges in this space, such as text recognition in low-resource languages, information extraction from historic/handwritten collections and multimodal modeling of complex document images. Additionally, we introduce new tasks that call for a top-down perspective— to understand a document image as a whole, not in parts—of document image understanding, different from the mainstream trend where the focus has been on solving various bottom-up tasks. Most of the existing tasks in Document Image Analysis (DIA) deal with independent bottom-up tasks that aim to get a machine-readable description of certain, pre-defined document elements at various abstractions such as text tokens or tables. This thesis motivates a purpose-driven DIA wherein a document image is analyzed dynamically, subject to a specific requirement set by a human user or an intelligent agent. We first consider the problem of making document images printed in low-resource languages machine-readable using an OCR and thereby making these documents AI-ready. To this end, we propose to use an end-to-end neural network model that can directly transcribe a word or line image from a document to corresponding Unicode transcription. We analyze how the proposed setup overcomes many challenges to text recognition of Indic languages. Results of our synthetic to real transfer learning experiments for text recognition demonstrate that models pre-trained on synthetic data and further fine-tuned on a portion of the real data perform as well as models trained purely on real data. For 10+ languages for which there have not been public datasets for printed text recognition, we introduce a new dataset that has more than one million word images in total. We further conduct an empirical study to compare different end-to-end neural network architectures for word and line recognition of printed text. Another significant contribution of this thesis is the introduction of new tasks that require a holistic understanding of document images. Different from existing tasks in Document Image Analysis (DIA) that attempt to solve independent bottom-up tasks, we motivate a top-down perspective of DIA that requires a holistic understanding of the image and purpose-driven information extraction. To this end, we propose two tasks—DocVQA and InfographicVQA— fashioned along Visual Question Answering (VQA) in computer vision. For DocVQA, we show results using multiple strong baselines that are adapted from existing models for existing VQA and QA problems. For InfographicVQA, we propose a transformer-based, BERT-like model that jointly models multimodal—vision, language, and layout—input. We conduct open challenges for both tasks, attracting hundreds of submissions so far. Next, we work on the problem of information extraction from a document image collection. Recognizing text from historical and/or handwritten manuscripts is a major challenge to information extraction from such collections. Similar to open-domain QA in NLP, we propose a new task in the context of document images that seek to get answers for natural language questions asked on collections of manuscripts. We propose a two-stage retrieval-based approach for the problem that

Abstract

Individual differences, encompassing a wide array of physical, cognitive, and emotional characteristics such as age, gender, personality, musical expertise, and empathy, play a pivotal role in shaping our interactions with the world around us. These differences, stable over time and across situations, not only influence our end goals but also the processes through which we perceive and interact with them. Recent research in neuroscience has brought to light that each individual possesses unique and fundamentally stable functional brain connections. These connections remain consistent, irrespective of the task at hand, suggesting that functional brain networks could potentially be employed as a measure of stable individual traits. Such an approach could revolutionize personalized medicine, offering tailored therapeutic interventions based on an individual’s unique brain signature. In the context of music, distinct patterns in functional brain networks related to individual differences become crucial. Music, as a continuous task, offers a naturalistic paradigm to explore these patterns. This study aims to bridge the gap in the existing literature by examining how brain responses to music may vary based on individual characteristics, such as gender and musical expertise. Furthermore, previous studies have predominantly focused on Pearson correlation, which captures linear relationships but may not encapsulate the full complexity of functional brain networks. Therefore, our study explores various functional connectivity (FC) measures. By comparing these measures, we aim to understand which ones are most suitable for a given study, ensuring that the chosen measures capture the intricacies of FC effectively. The study utilized data from the ”Tunteet” project, which involved multiple fMRI scans and behavioral tests. A total of 36 participants underwent fMRI scanning while listening to three distinct 8-minute-long musical pieces representing different musical styles. Their responses were analyzed using various temporal and spectral FC measures. The aim was to compare these FC measures to identify the one that captured the most variation associated with gender and musical expertise. Subsequently, a binary Support Vector Machine (SVM) was employed to classify distinct population groupings. Our results align with previous research, suggesting that musical preferences can indeed be considered as a distinct personality trait. This was particularly evident in the differences in liking ratings specific to gender and musical expertise. We also found that Coherence, a spectral-domain FC measure, captured the maximum variation for musical stimuli. However, when classifying individuals based on gender and musical expertise, a composite measure, which combined all measures obtained by concatenation followed by a feature selection procedure, outperformed any single measure. This composite measure consistently achieved better results, suggesting that each FC measure captures different aspects of the relationship between brain regions. In summary, our research offers a fresh perspective on the interplay between musical preferences, gender, musical expertise, and brain responses. By leveraging diverse functional connectivity measures, we’ve shed light on the complexities of functional brain networks, paving the way for future research in this domain.

Abstract

As the number of applications of Internet-of-things (IoT) is increasing daily, there is a growing challenge of providing real-time tracking data to users with a larger coverage area. The satellite-based network can provide global coverage to IoT devices. Due to the low power and low complexity of IoT devices using slotted-ALOHA protocol, it is desired to have visible satellites at a lower altitude than that altitude of commonly used GEO satellite constellations. Recently, sizeable Low-earth-orbit (LEO) satellite constellations such as Starlink-SpaceX and OneWeb launched in orbit can address the coverage issues at much lower power and lower latency. In addition to this, there is minimal coordination among IoT devices. Thus, the packets from multiple devices reaching the gateway may cause collisions and thus reduce the throughput at the gateway. This thesis considers a LEO satellite-based topology for an IoT network, where multiple IoT devices broadcast the information to all the visible satellites over a shared channel using slotted ALOHA. The satellites selectively Decode-and-forward (DF) the data from the IoT devices over orthogonal channels to the Ground station (GS), which does Maximal Ratio Combining (MRC). Capture and Successive interference cancellation (SIC) schemes are considered for decoding at the satellites to mitigate the interference at the satellites in the uplink. For the considered topology, the closed-form expressions are derived for the end-to-end Outage probability (OP) for an arbitrary number of IoT devices and satellites in the capture model and the two-device, two-satellite case in the case of the SIC model. The expressions are derived for both independent and non-identically distributed (inid) and independent and identically distributed (iid) uplink channels. The OP is analyzed as a function of the parameters like the number of satellites, devices, and the desired data rate. The results demonstrate that the proposed approach leverages the benefits of mega-LEO satellites to make the topology feasible and attractive for low-powered and low-complexity IoT networks.

Abstract

Water resources sources have deteriorated in recent years due to rapid population growth, chemical industry effluents, agricultural practices, and climate change. The primary cause of contamination of water-bodies is due to disposal of untreated sewage water pollution. Hence effective sewage treatment is necessary to safeguard water-bodies from contamination. Many cutting-edge techniques have been developed in recent years to increase the effectiveness of the removal of organic matter and nutrients by Waste Water Treatment Plants (WWTPs). To comprehend the effectiveness of wastewater treatment, the current study considered a Sewage Treatment Plant (STP) located in Amberpet, Hyderabad. The capacity of the STP at Amberpet is 339 MLD (Million Liters per Day) and is evaluated by collecting 156 samples for 12 months (January 2018 – December 2018). The data collected was based on grab and composite collection. In STPs, grab sampling and composite sampling are two frequently used techniques for gathering wastewater samples. Grab sampling is the process of taking one wastewater sample at a specific time. Contrarily, composite sampling involves collecting multiple effluent samples over a predetermined period. An STP aims to minimize or remove organic debris, sediments, disease-causing organisms, and other pollutants in sewage water before disposing into streams and other water bodies. The current study observed the removal efficiencies of the constituents such as Total Suspended Solids (TSS), Biological Oxygen Demand (BOD), Chemical Oxygen Demand (COD), and several other parameters. This investigation assesses whether the effluents emitted into the river body are within the National River Conservative Directorate (NRCD) set limitations as the treated sewage water is discharged into the Musi River (a tributary of Krishna Basin). This study investigated the seasonal variation of WWTP efficiencies for different water quality parameters for the year 2018. The study was conducted over four seasons: winter (December, January, February), pre-monsoon (March, April, May), monsoon (June, July, August, September), and post-monsoon (October, November). The study analyzed the influents and effluents of each available water quality parameter, using both grab and composite sampling and the performance efficiency of STP. For each water quality parameter, the findings revealed considerable seasonal fluctuations in treatment plant efficiencies. For example, if Dissolved Oxygen (DO) level is too low, the microorganisms may become stressed or die, leading to reduced treatment efficiency. On the other hand, if DO level is too high, it can lead to the growth of aerobic bacteria that consume DO, reducing the available oxygen for the treatment process. Therefore, it is important to monitor the DO levels in the STP regularly and maintain them within the recommended range to ensure efficient and effective treatment of the wastewater. The type of treatment technique used in STP determines the DO saturation level necessary for a sewage treatment facility. For an activated sludge process, the DO saturation level should typically be kept between 2 and 3 mg/L, and between 1 and 2 mg/L for a trickling filter process. Water temperature is a prominent variable for water quality and aquatic habitat affecting saturation dissolved oxygen concentrations, algal metabolism, fish growth, and production in aquatic systems. Water temperature also signifies the health of the river water body and regulates many physical and chemical parameters related to river water quality parameters, which speculatively depend on many factors. The study provided valuable insights into the influence of water temperature on saturation oxygen levels in the water, highlighting the importance of considering temperature as a crucial environmental factor in assessing water quality. So far, most River Water Temperature (RWT) models are either physical or datadriven models requiring large amounts of hydrological and meteorological observations. Many climate change studies have been conducted with increasing stream water temperatures, but how it affects saturated DO levels have not been addressed. To address these, the present work aims to work with a hybrid model - Air2Stream as a function of air temperature and discharge and also demonstrates how the Air2Stream method can be used to generate accurate RWT predictions and subsequent DO concentrations in river water quality modeling. RWT, which combines the ideas of the heat budget equation that generalizes physical processes and infers relationships between input and output data, can be predicted using the Air2Stream model. With two river gauging stations at Mantralayam, Shimoga in the Krishna River Basin, the efficiency of the suggested modeling framework is demonstrated. By comparing simulated RWT time series with matched observations and calculating the Nash Sutcliffe Efficiency (NSE),

Abstract

Sentiment Analysis is an important task for analysing online content across languages for tasks such as content moderation and opinion mining. However, state-of-the-art NLP modelling techniques often require a large amount of training data to achieve their results. Unfortunately, high-quality annotated data is often a rare commodity for many languages other than English, including most Indian languages. We attempt to tackle this data scarcity in this thesis in two ways - by creating additional resources, and by exploring more data-efficient modelling techniques. Over the past few years, while some significant resources for Sentiment Analysis have been developed in several Indian languages, there do not exist any large-scale, open-access corpora for Gujarati. In this thesis, we present and describe the Gujarati Sentiment Analysis Corpus (GSAC), which has been sourced from Twitter and manually annotated by native speakers of the language. We describe in detail our collection and annotation processes and conduct extensive experiments on our corpus to provide reliable baselines for future work using our dataset. We then explore modelling techniques that work well in a low-resource setting by experimenting with AfriSenti, a collection of sentiment analysis datasets in 12 African languages. We propose an XGBoost-based ensemble model trained on emoticon frequency-based features and the predictions of several statistical models such as SVMs, Logistic Regression, Random Forests, and BERT-based pretrained language models such as AfriBERTa and AfroXLMR. We also report results from additional experiments not in the system and conduct an ablation study to observe the effects of different types of models and features on the final ensemble.

Abstract

The human brain is a highly complex network of interconnected neurons exhibiting emergent phenomena. The theory of the brain functioning at a critical state - near a second-order phase transition where its capabilities for memory, learning, and information processing - has gained widespread popularity in recent years. This criticality has often been likened to the critical state observed in the Ising model, which is influenced by a temperature parameter. Neural dysfunction due to neurodegenerative conditions such as Alzheimer’s Disease influences the proximity to this critical state. We study the disrupted proximity to criticality of the dysfunctional Alzheimer’s brain using a spin-lattice pairwise maximum entropy model where connectivity is inferred from resting-state fMRI data. The closeness to criticality is quantified through the critical temperature, which is derived by analysing the susceptibility of the system. We find an insignificant difference between the two classes, which we attribute to the dependence of susceptibility on spatial dynamics, with the hypothesis that the existence of a spin-glass-like criticality could hinder the sensitivity of this metric. To characterise the nature of brain criticality, we explore the spatial characteristics and temporal correlations of criticality for both normal and Alzheimer’s brains. We find a weak distinction in the spatial dynamics between the two classes - the sizes and number of spatial domains formed, as well as the self-averaging behaviour of the system, were not found to vary significantly. However, we notice a stretched exponential form in the temporal relaxation times, which is characteristic of spin-glasses. We find a significant difference between the relaxation behaviour of Alzheimer’s and Cognitively Normal subjects. In spin-glass systems, a third phase exists near the phase transition boundary - the spin-glass phase. This phase is characterised by frustration and competing interactions. It allows for the emergence of complex properties such as long-term memory, which are features not found in simple Ising-like criticality. Despite the difference in closeness to criticality, the lack of a significant distinction in the spatial correlation lengths can be attributed to local frustrations, which might not allow the system to reach equilibrium domain distributions. Based on the differences observed between the spatial and temporal correlations, we conclude that the criticality found in the brain is more akin to a spin-glass criticality, which is more sensitive to temporal rather than spatial measures.

Abstract

Education plays a vital role in shaping up one’s life. Education in early childhood includes learning from different activities where counting and other concepts of mathematics act as major building blocks. Mathematics is not just a subject to be taught in schools, but also it has myriad applications in our daily lives such as counting the total number of stationery items, calculating their individual prices, and adding them up for a purchase made in a grocery shop. This kind of real world situations are posited in mathematical word problems which are an essential part of a child’s learning that requires natural language understanding (NLU) as well as knowledge of mathematical operations. Mathematical Word Problem Solvers can assist both students and teachers. It is a challenging field and this thesis attempts to provide NLP solutions for math word problems in English and Indian languages. Our approach for developing word problem solvers follows a pipeline. For a word problem, first, we identify the relevant operands and required operations. In the second step, we form an equation from these identified components. The first two stages can be combined to generate equations at once using neural network based approaches. At the last stage, the equation is solved by a mathematical solver to get the final solution. We focus primarily on the first two stages of this pipeline. We developed solvers using three kinds of approaches: frame based, composition of classifiers based, and end-to-end neural based. We also shed light on the limitations of the current automatic solvers with respect to the data. We designed different data augmentation techniques to overcome the data scarcity problem. As a part of resource building, we developed word problem datasets and solvers for Indian Languages. In addition to this, we compared different models related to our developed approaches. We empirically show the difference in generation of various equation notation types. For this study, we present the results of equations in infix, postfix, and prefix notations. We also show two natural language processing tasks where components of word problem solvers can be utilized. In the first task, we studied the impact of simple number based pre-processing on the performance of machine translation systems. In the second task, we analyze speech transcript texts to extract equation spans. This kind of text is present mainly in transcripts from mathematical domain. For achieving this, we develop an equation identifier and convertor involving math notations for transcripts. This can make the transcripts easily readable for transcripts which have wide usage of mathematical terms.