Abstract

Integrating Artificial Intelligence (AI) into drug design has revolutionized the field at various levels, ranging from simple molecular property prediction tasks to complex de novo drug design tasks. This is still a relatively young field and is rapidly evolving. In this thesis, we identify various challenges AI models face due to data constraints and analyze hidden biases in these methods; we improve upon existing molecular generation models for aligning the drug design to desired complex properties; we propose a novel framework for structure elucidation from molecular spectra. The second chapter explores various biases in protein-ligand datasets that affect the performance of deep learning models in binding affinity prediction and virtual screening. It demonstrates how random data splitting leads to overly optimistic results and introduces more nuanced data splitting methods that account for sequence, pocket structure, and protein-ligand interaction similarities. The research identifies biases in these datasets related to protein-only and ligand-only information. It proposes improvements in dataset construction and model design to enhance the generalizability and accuracy of AI-based protein-ligand scoring functions. In the third chapter, we develop and refine Molecular generative models for more accurate conditional generation. We propose optimization methods aligning the generated molecules’ properties with the desired ones. The fourth chapter introduces the Spectra and Molecule Encoder Network (SMEN) for scoring molecules against target spectra. The model aims to aid library ranking tasks for Infrared (IR) Spectrumbased structure elucidation. SMEN learns spectral and molecular embeddings, accurately representing both entities in a high-dimensional latent space. Additionally, the SMILES Decoder (SD) generates SMILES strings from the latent space, enabling accurate one-shot predictions from IR spectra to molecule structures and showcasing promising accuracies for rapid, precise structure elucidation. In summary, This thesis explores the potential of ML models to predict binding affinities and generate molecular structures with desired properties. It begins by identifying and addressing latent biases in popular protein-ligand datasets, affecting binding affinity predictions’ accuracy. This research introduces more refined data-splitting methods that enhance model generalizability by analyzing biases in sequence-based and structure-based datasets. The thesis then presents a generative model for generating molecules with specific characteristics, such as binding affinity, and further optimizes the model’s ability to produce molecules that meet predefined criteria. Lastly, a contrastive learning framework is developed for generating molecular structures from infrared spectra, demonstrating the model’s versatility in handling various molecular generation tasks. Each chapter addresses foundational issues, setting the stage for future innovations in the field.

Abstract

This research delves into the application of advanced semantic segmentation techniques to automate facade inspection, a critical component in the maintenance and safety assessment of urban infrastruc- tures. Central to the study is the WALL-E dataset, a comprehensive collection specifically curated to encompass a wide range of facade defects, including cracks, stains, and delamination. This dataset is designed to challenge and refine the capabilities of deep learning models by exposing them to a diverse array of defect types and complex architectural features, thereby mimicking the variety of conditions encountered in real-world settings. The training methodology employed in this study leverages the strengths of the DeepLabV3+ model, compared against other leading models such as UNet, PSPNet, and FCN. Each model was rigorously trained and evaluated on their ability to accurately segment defects across different types of building materials and architectural styles. DeepLabV3+ emerged as the standout model due to its sophisticated architecture that combines atrous convolution with an encoder-decoder structure, enhancing its ability to capture detailed spatial information and accurately delineate defect boundaries. A significant portion of the evaluation focused on the generalization capabilities of the models. DeepLabV3+ demonstrated remarkable adaptability, not only performing well on the diverse images of the WALL-E dataset but also showing robust generalization on arbitrary internet images and out-of- distribution samples. This indicates the model’s ability to extend beyond its training parameters and adapt to new, unseen scenarios, which is essential for practical applications where variable conditions are common. Further assessments were made on images without defects to evaluate the model’s precision in iden- tifying defect-free areas, thereby reducing the likelihood of false positives. This aspect of the model’s performance is critical in practical scenarios, where the ability to distinguish between damaged and intact structures can significantly impact maintenance decisions and safety assessments. This thesis highlights the significant potential of using semantic segmentation in the realm of struc- tural health monitoring. The advanced capabilities of DeepLabV3+, coupled with the comprehensive and diverse WALL-E dataset, set a new benchmark in the field, providing a strong foundation for future research and practical applications in facade inspection. This work not only advances the technology in automated building inspections but also opens up new avenues for further enhancements and applica- tions in urban infrastructure maintenance.

Abstract

Federated learning offers a promising approach to learn collaboratively across multiple devices or agents. This collaboration allows leveraging a vast amount of data while preserving privacy by keeping the raw data local. However, this approach is not void of challenges. This thesis explores some of these challenges in the context of multi-armed bandit (MAB) problems. MAB problems are a type of decision-making problem where an agent repeatedly chooses between multiple actions, each with an unknown reward, and aims to maximize the total reward over time. Federated MABs introduce unique fairness concerns. For instances, for a fair algorithm, the learning process should not favor certain actions over others to an extreme degree, especially if those actions are beneficial to specific sub-populations. In addition to fairness, privacy is another key challenge. While federated learning avoids sharing raw data, the learning process itself can leak information about individual data points. This thesis proposes two novel algorithms (and variants) to address these challenges. The first algorithm, P-FCB, tackles a specific type of MAB problem with constraints in a federated setting. It promotes fairness by optimizing for the collective benefit of all users. The second algorithm, Fed-FairX-LinUCB, focuses on achieving fairness and privacy guarantees in a more complex scenario where actions have additional context. It ensures that action selection is fair and avoids situations where some actions are never chosen. Additionally, both these algorithms provide differential privacy guarantees, a formal guarantee that ensures the learning process reveals minimal information about any individual data point. By exploring these algorithms, this thesis aims to contribute to the development of federated learning for MAB problems that is both fair and privacy-preserving.

Abstract

Sleep plays a critical role in maintaining overall health and well-being, impacting both psychological and physiological functions. However, sleep disorders such as insomnia, obstructive sleep apnea (OSA), and narcolepsy are becoming increasingly prevalent, significantly affecting individual’s quality of life. Effective sleep stage classification is essential for diagnosing and treating these disorders. Traditional methods like polysomnography (PSG) are the gold standard for diagnosing sleep disorders, but they are expensive, cumbersome, and not easily accessible due to the need for specialized equipment and trained personnel. Moreover, PSG relies heavily on electroencephalogram (EEG) data, which, despite its effectiveness, presents practical challenges in non-clinical settings due to the complexity of electrode placement and patient discomfort. Innovative strategies in wearable technology using supervised deep learning techniques offer a promising alternative for enhancing sleep stage classification, making it more accessible and user-friendly. This work aims to overcome these limitations by exploring the potential of using electrooculogram (EOG) signals for sleep stage classification. EOG signals, which are non-invasive and can be more easily integrated into wearable technology, offer a promising alternative for long-term sleep monitoring and provide a more comfortable experience for patients. The core contribution of this research is the design and implementation of a novel SE-ResNet-Transformer model. This model architecture combines the strengths of Squeeze-and-Excitation Residual Networks (SE-ResNet) and Transformer encoders, tailored to handle the unique characteristics of EOG signals. The SE-ResNet component enhances feature extraction by adaptively recalibrating feature responses, allowing the model to focus on the most informative aspects of the EOG signal. The Transformer encoder component captures the temporal dependencies within the EOG signals, enabling the model to understand the dynamic transitions between sleep stages. The performance of the SE-ResNet-Transformer model was rigorously evaluated using three publicly available datasets—SleepEDF-20, SleepEDF-78, and SHHS. The model demonstrated superior accuracy and robustness across different datasets, with its EOG-based performance comparable to models trained with EEG data. This suggests that EOG is a viable alternative for sleep monitoring, particularly relevant for the development of wearable technology. Extensive ablation studies were conducted to fine-tune the model’s parameters and configurations, providing valuable insights into optimal settings for window sizes and strides, and enhancing the model’s efficiency and performance. Techniques like 1D-Gradient-weighted Class Activation Mapping (Grad-CAM) and t-distributed Stochastic Neighbor Embedding (t-SNE) plots were employed to visualize the model’s decision-making process, emphasizing transparency and interpretability. In addition, this work introduces the pioneering Indian SLEEP dataset of ischemic Stroke patients (iSLEEPS) which addresses a gap in the availability of India-specific sleep disorder data, with a particular focus on the impact of ischemic stroke on sleep health. This dataset is comprehensive, including detailed clinical annotations that cover patient demographics, diagnoses, and sleep stages. Cutting-edge deep learning models, including CNN-based, LSTM-based, and Transformer-based architectures, were trained on raw single-channel EEG and EOG signals. Among these, the Transformer-based model showed remarkable performance in sleep stage classificatzions using single-channel EOG data. The successful validation of these models on the iSLEEPS highlights their potential to enhance diagnostic tools and therapeutic strategies, thereby contributing to improved patient outcomes and fostering global collaboration in sleep research. In conclusion, this thesis presents advancements in the field of automated sleep stage classification using wearable technology and supervised deep learning. By leveraging EOG signals and deep learning architectures, this research offers promising solutions for more accessible, accurate, and comfortable sleep monitoring, ultimately aiming to bridge the gap between clinicalgrade sleep monitoring and everyday health management tools, thus enhancing sleep health and overall wellness

Abstract

The strategic application of artificial intelligence (AI) in healthcare has the potential to revolutionize patient care, particularly in the domains of public health and cardiology. This thesis presents innovative AI-based solutions to address critical challenges in these fields, focusing on enhancing decision-making, disease surveillance, and diagnostic accuracy. In the second chapter, we introduce the AI-Based Home-Based Care Assistive System (AI-HBCAS), developed to aid healthcare professionals in identifying severe patient conditions that require referrals. By analyzing patient demographics, symptoms, and vital signs, AI-HBCAS enhances clinical decision- making, ensuring timely and appropriate care for patients with non-communicable diseases (NCDs). This system aligns with the Ministry of Health and Family Welfare’s goals of improving public health delivery through digital health solutions. The third chapter explores the development of a Pharmaco- surveillance model for outbreak detection and awareness in Primary Health Centers (PHCs). Leveraging AI, we try to analyse prescriptions and pharmacy records to map drug utilization patterns and detect lo- cal outbreaks. The insights generated could potentially support timely public health interventions and contribute to the Ministry of Health and Family Welfare’s objectives of enhancing public health infras- tructure and outcomes. We then delve into the development of a unique deep learning architecture for heart murmur detection using phonocardiogram (PCG) recordings in chapter four. Our modified variable kernel length Resid- ual Networks (ResNets) demonstrate significant improvements in detecting heart murmurs, achieving a weighted accuracy score of 0.708 in the George B. Moody PhysioNet Challenge 2022. This inno- vative approach highlights the potential of deep learning models in improving diagnostic accuracy in cardiology, specifically using audio modalities. Chapter five focuses on the creation of a machine learn- ing approach for predicting outcomes in post-anoxic coma patients using longitudinal EEG and ECG recordings. Our model extracts frequency domain features from these recordings, combining them with demographic data to predict neurological recovery following cardiac arrest. This solution, part of the George B. Moody PhysioNet Challenge 2023, underscores the viability of machine learning models in critical care settings, offering valuable prognostic insights.

Abstract

This thesis explores and analyzes humor generation through several methods. This thesis introduces an approach to automated joke generation that leverages a novel hybrid model combining template extraction and infilling. Recognizing the inherent structure and unpredictability in humor, this work pi- oneers a method that extracts templates from existing jokes and subsequently infills these templates with new content, guided by state-of-the-art (SOTA) techniques available at the time of research. Specifically, the thesis accomplishes three main objectives: Firstly, it develops an automated system for joke genera- tion that innovatively combines template extraction with a deep learning approach for content infilling, ensuring both the preservation of joke structure and the introduction of novelty. This approach uses a rule-based method to extract a template from a joke and then infills it using BERT. Secondly, it extends the base model by incorporating advancements in language models and natural language processing to enhance the flexibility and quality of generated jokes, broadening the scope of humor that the system can produce. This approach uses a novel strategy to identify semantically salient words, and uses LLMs on top of earlier models for the infilling. Lastly, the thesis explores a distinct avenue in humor genera- tion by employing conversational prompts within a conversational AI framework with GPT-4, the SOTA at the time of presentation. This method not only diversifies the ways in which humor can be generated but also reflects on the potential of conversational AI to contribute creatively to the domain of humor. Through rigorous evaluation, this work demonstrates the efficacy of the proposed methods in generating humor that is coherent, contextually appropriate, and, most importantly, funny. This research not only contributes a novel technique to the field of computational humor but also opens up new pathways for future exploration in automated joke generation and the creative capacities of AI.

Abstract

Water, a ubiquitous molecule in life and chemical processes, undergoes autoionization, where it splits into a proton (H+) and a hydroxide ion (OH-). This process plays a crucial role in determining pH and conductivity. However, the mechanism of autoionization remains a mystery due to the vast difference in stability between water and dissociated ions. This difference creates a multiscale kinetic challenge, making it difficult to simulate the entire process using traditional methods. This thesis investigates the water auto-ionization mechanism using well-tempered metadynamics, a powerful tool for exploring free energy landscapes. We employ atomic neural network potentials (NNPs) to overcome the computational limitations of traditional density functional theory (DFT) calculations, allowing us to explore the ionization pathway at a nanosecond timescale. Central to our methodology is the identification of an effective collective variable to drive the ionization in a way that doesn’t break the dynamics of water or disturb the free energy surface. Our approach of using a group of water molecules in vicinity and biasing the molecules to align to a favourable configuration for ionization allows us to initiate ionization on our choice of water molecules. In our thesis we demonstrate the evolution of our collective variable arising from the observed mechanics of auto-ionization along with the various observed patterns in which the ions exist in the system. We have employed a few algorithms to identify the existence of ions. The combination of our collective variables and the analysis algorithms, gives us a more definite molecular understanding of the dynamics defined in previous related work. We demonstrate that our observed dynamics are in excellent agreement with the physical qualities of water and also well represent the dynamics of ionization talked about in previous articles. With the collected data we also attempt to conduct importance classification of various parameters that might describe the initiation conditions of ionization with a good accuracy. In the bigger picture, this understanding may allow us to get closer towards using machine learning neural network potentials to reliably study the dynamics of other water implicit systems and hence gather more information about auto-ionized water and its effects on such systems. In essence this thesis demonstrates the use of neural network potential n2p2 as a reliable method of simulating ions in water without disturbing the dynamics of water or the ions. The use of neural network potential instead of DFT calculations allows us to study simulations with significantly longer timescales within the same processing constraints as before. This can be a significant step towards studying systems that are affected by the auto-ionization of water eg. water implicit protein folding.

Abstract

Sleep staging using Electroencephalogram (EEG) signals is crucial for assessing sleep quality and diagnosing related disorders. This thesis explores methodologies for automated sleep stage classification leveraging EEG data, focusing on both supervised learning and self-supervised learning (mulEEG). The mulEEG study introduces a novel multi-view self-supervised method designed for EEG representation learning. The mulEEG framework utilizes complementary information from multiple views to learn robust representations without requiring labeled data. Key contributions of mulEEG include an EEG augmentation strategy tailored for multi-view self-supervised learning, the introduction of a diverse loss function to enhance complementary information across views, and demonstrating the superiority of mulEEG over traditional supervised methods. The framework achieves this by combining different EEG augmentations such as jittering, flipping, and scaling, and processing these augmentations through both time-series and spectrogram encoders to capture diverse features. This joint training strategy, coupled with a unique contrastive loss function, significantly improves representation learning, making mulEEG highly effective for tasks like sleep staging. Results indicate that the mulEEG method outperforms supervised models in transfer learning scenarios, showing improvements of 1.1% in Cohen’s kappa and 0.85% in accuracy compared to traditional supervised learning models. In a second contribution, we explored supervised learning strategy for EEG sleep stage classification integrating squeeze and excitation (SE) blocks within a residual network and stacked Bidirectional Long Short-Term Memory (Bi-LSTM) units. This model aims to capture complex temporal dependencies and spatial features in EEG data. A significant aspect of this strategy is the application of GradCam for model interpretability, allowing visualization of the model’s decision-making process and allows comparison with sleep experts’ insights. The model was evaluated on several publicly available datasets (SleepEDF-20, SleepEDF-78, and SHHS), achieving high Macro-F1 scores of 82.5, 78.9, and 81.9, respectively. This demonstrates the model’s robustness and applicability across different datasets. Additionally, the study introduces a training efficiency enhancement strategy, reducing training times by eight-fold with minimal impact on performance. In conclusion, these EEG-based sleep staging approaches hold promise for enhancing sleep quality assessment and facilitating the diagnosis of sleep-related disorders, significantly contributing to the field of biomedical informatics. The incorporation of model interpretability provides valuable insights into the underlying decision-making processes, effectively bridging the gap between machine learning models and clinical practice.

Abstract

The simultaneous occurrence of climate extremes poses significant threats to ecosystems, amplifying the vulnerability to various physical risks. A compound extreme event corresponds to a situation in which multiple extremes occur at the same time, leading to more severe impacts than individual extremes. Despite extensive research on individual hot extremes and droughts globally, and the development of numerous indicators to detect and study changes in extreme events across space and time, there remains a substantial gap in understanding the occurrence, magnitude, spatial extent, and associated risks of compound extremes. Most current climate extreme indicators do not provide information on compound or concurrent events. These compound extremes encompass scenarios such as warm/dry, cold/dry, warm/wet, and cold/wet conditions. Comprehensive research integrating multiple climate variables is crucial to address this gap, providing a holistic understanding of the risks and impacts associated with compound extremes. This study aims to fill this gap by investigating various compound extreme scenarios through the analysis of combinations of maximum temperature (Tx) and the Standardized Precipitation Evapotranspiration Index (SPEI). The analysis uses monthly data from 1951 to 2014, obtained from the India Meteorological Department (IMD) for the Indian landmass. The results of the study reveal several key trends and patterns in the spatial extent and frequency of compound extreme events. Notably, the spatial extent of warm/dry events has increased at a rate of 1.8% per decade, indicating a significant upward trend. Conversely, cold/wet events have shown a decrease of 1.1% per decade, highlighting a reduction in these events over the Indian region. Warm/wet events have also exhibited an increasing trend, albeit at a more modest rate of 0.3% per decade, while cold/dry events have shown a modest rise of 0.7% per decade. Furthermore, the study examines the frequency and return periods of these compound extreme events. It is observed that compound warm/dry and cold/wet extremes exhibit high frequency and shorter return periods, posing greater risks to the affected regions. On the other hand, compound cold/dry and warm/wet extremes occur less frequently, indicating longer return periods and consequently lower associated risks. The frequency analysis reveals that across much of the country, the occurrence of warm/dry, cold/dry, warm/wet, and cold/wet extremes ranges from 30-45 months, 15-30 months, 20-30 months, and 25-45 months, respectively. A notable finding of the study is the increased frequency of warm/dry conditions in the recent period (1983-2014) compared to the base period (1951-1982). Specifically, the recent period experienced approximately 31 years of warm/dry conditions exceeding a spatial extent of 5%, whereas the base period had approximately 24 years. This trend underscores the growing prevalence of warm/dry extremes in recent decades. The findings of this study contribute to an enhanced understanding of the changes in compound climate extremes from a multivariate perspective. By examining the interplay between temperature and precipitation extremes, this research provides valuable insights into the evolving nature of climate extremes in India. The observed trends in the spatial extent and frequency of compound extremes have significant implications for climate risk assessment and adaptation strategies. The increasing trend of warm/dry events suggests a heightened risk of heatwaves and droughts, which can have severe impacts on agriculture, water resources, and human health. The study highlights the need for a comprehensive understanding of compound climate extremes to inform effective adaptation and mitigation strategies. Policymakers and stakeholders can leverage the findings of this research to develop targeted measures to enhance resilience against compound extremes. For instance, the increasing frequency of warm/dry conditions calls for improved water management practices, drought-resistant crop varieties, and heatwave preparedness plans. Moreover, the study emphasizes the importance of considering compound extremes in climate models and projections. Traditional approaches that focus on single variables may underestimate the risks associated with simultaneous extreme events. By incorporating compound extremes into climate models, researchers can better inform decision-making processes.

Abstract

In the realm of Natural Language Processing (NLP), enhancing knowledge graphs (KGs) with large language models (LLMs) offers a transformative approach to managing and leveraging structured information. This thesis, titled ”Leveraging Large Language Models for Knowledge Graph Enhancement: From Creation to Utilization,” explores the synergistic integration of LLMs and KGs to enhance various aspects of information extraction, entity classification, and question answering. The research is driven by the recognition of the limitations inherent in traditional KGs, which often suffer from incomplete but dynamic data, and aims to leverage the semantic capabilities of static LMs to address these gaps. The thesis presents a comprehensive study on cross-lingual fact extraction, focusing on extracting factual information from texts in low-resource Indian languages and representing it as English triples. This task is critical for capturing the rich knowledge embedded in these languages, which traditional methods often overlook. An end-to-end generative approach is employed, achieving a significant F1 score of 77.46, thereby setting a new benchmark for cross-lingual fact extraction efforts. This method is validated using the XAlign dataset, which includes multilingual data from Wikipedia articles, ensuring robust evaluation and applicability of the proposed models. Entity type classification is another focal point of this thesis. The research introduces a fine-grained entity type classification model, which is particularly effective in scenarios where KGs are being developed from scratch using only textual descriptions. This model, built upon a multilingual encoder- only language model and incorporating hierarchy-aware losses, utilizes our Wikidata Entity Type Classification (WETC) dataset. By leveraging the structure of Wikidata and the multilingual links in Wikipedia articles, the model achieves significant improvements in entity classification across various languages, attaining a performance of 0.73 Hierarchical F1. The thesis also delves into the integration of KGs with LLMs for question answering (QA). A novel approach, termed Joint Reasoning over KG and LLMs, is proposed, where the structured knowledge from KGs is combined with the generative abilities of LLMs. This hybrid architecture employs graph neural networks (GNNs) and includes modifications such as Prominent Entities for Graph Augmentation (PEGA) and QA-Context-Aware Node Pruning (CANP) to enhance the relevance and accuracy of the QA process.Experimental results demonstrate that this integrated approach significantly outperforms existing methods on various QA benchmarks, including OpenBookQA, CommonsenseQA, and MedQA-USMLE, showing notable advancements in QA accuracy over 20% points. In conclusion, this thesis makes substantial contributions to the field of knowledge graph enhancement using large language models. By addressing the challenges associated with cross-lingual fact extraction,entity classification, and QA, the research paves the way for the development of more advanced and efficient knowledge-based systems by bridging the gap between structured knowledge representation and natural language understanding.