Abstract

Dementia is a chronic and progressive syndrome that affects the cognitive functioning of an individual. Alzheimer�s, a neurodegenerative disorder, is the leading cause of dementia. The only way to control the progression of the disease is its early detection, followed by drug and non-drug interventions. The speech production chain is the presentation of cognitive abilities and is majorly affected in the early stages of Alzheimer�s disease. Speech signals are ubiquitous and facilitate easy recording, storage, and transfer. For these reasons, researchers have long strived to develop complementary tools for Alzheimer�s Dementia detection using acoustic and linguistic clues derived from speech utterances. This thesis is one more attempt at finding the differentiating auditory and linguistic clues in the utterance(audio and transcript) of an Alzheimer�s Dementia patient. The ADReSS INTERSPEECH-2020 and ADReSSo INTERSPEECH-2021 challenges provide a balanced dataset for the Alzheimer�s Dementia classification task. This thesis explores the efficacy of different acoustic features to capture distinct patterns in the speech utterance of AD patients. The features are obtained by evaluating Cepstral Coefficients over different acoustic algorithms and techniques. Mel-frequency and Linear Prediction methods are used to capture the Vocal tract characteristics; Residual Coefficient, Glottal Volume Velocity, and Zero Frequency Filtering approach for excitation source characteristics; Envelope Modulation Spectrum and Long Term Averaging Spectrum capture the prosody characteristics of speech utterance. The next part of this thesis explores the Single-frequency-filtering-based(SFF) high spectrotemporal resolution feature using the filter bank approach for Alzheimer�s Dementia detection. Experiments are performed using different machine learning classifiers over the acoustic features extracted from the challenge datasets. The current study also demonstrates the performance of the BERT model for the dementia classification task. Finally, the performance of individual and combined acoustic features is reported. Also, the classification score is evaluated by score level fusion of the acoustic models and BERT model to observe the complementary characteristics of acoustic features to the BERT model. Acoustic models perform best when combined with linguistic features, suggesting the complementary nature of acoustic and linguistic features. Also, the high spectro-temporal resolution Single Frequency Filtering feature captures the characteristics of speech patterns better for Alzheimer�s Dementia classification than traditional source-filter model-based features.

Abstract

Discovering new reaction pathways lies at the heart of drug discovery and chemical experimentation. Chemical patent texts contain new reactions and reaction pathways. Thus, a huge amount of drug reaction data lies in unannotated patent texts which are not machine-readable. Reaction roles play an important part in analysing chemical pathways, and tracing chemicals through them, and while there is a vast body of chemical data available, the unavailability of reaction role annotated data is a blocker to effectively deploy deep learning methods for reaction discovery. To overcome this hurdle, this work introduces a new dataset, WEAVE 2.0, an expansion of the existing WEAVE dataset, a chemical NER dataset obtained from chemical patents. WEAVE 2.0 augments WEAVE named entities along with full, manual, annotations of novel chemical reactions with reaction role information. We provide baseline models for chemical entity recognition from our raw dataset, using simple architectures commonly used for chemical NER and related tasks, such as biomedical NER. As the size of the dataset is small, we introduce dataset augmentation techniques to improve learning. These techniques can be used to generate further data from other patent-based datasets, such as WEAVE [17]. We also introduce and test improved models, which structure the problem as two smaller parts instead of one, both against the raw dataset, as well as data augmented using the above methods. Further, we compare our best models against other similar datasets for chemical NER, showing its performance across multiple similar tasks. Our dataset and associated models form the foundation of neural understanding of chemical reaction pathways via reaction roles and will allow models trained for downstream tasks to utilise this information to generally lead to better predictions.

Abstract

In the current digital era, there has been an exponential surge in the demand for data storage and retrieval. This escalating need is propelled by a diverse range of applications, spanning from cloud computing and data centers to mobile devices and embedded systems. NAND flash memory, a non-volatile storage technology, plays a pivotal role in meeting these demands and is extensively employed in solid-state drives (SSDs), memory cards, and various other applications. Its notable feature is its high storage density, enabling the storage of substantial volumes of data in a compact and efficient form factor. Industries have recently favored 3D NAND Flash over Traditional NAND Flash, where the �3D� denotes the vertical stacking of memory cells in multiple layers, akin to a skyscraper, to amplify storage density. While this stacking brings benefits, it presents challenges for analog circuit design engineers. As the number of layers increases, so does the load capacitance, introducing complexities in design as each new generation witnesses a rise in Load Capacitance. This challenge persists despite the growing number of vertical layers, particularly affecting the development of a robust and stable Low Dropout Regulator (LDO) amidst the increasing load capacitance in each generation. Moreover, the escalating layer count necessitates a proportionate increase in the number of LDOs to adeptly supply power to this expansive distributed network. To tackle the aforementioned challenges, the thesis introduces a novel LDO featuring a Flipped Voltage Follower based driver (FVF), deviating from conventional drivers. This innovative design offers increased robustness, especially when applied to a 3D NAND Flash memory array, surpassing the capabilities of traditional architectures. The thesis advocates for a dualloop scheme to enhance the transient performance metrics of the LDO. This proposed scheme not only diminishes settling time but also reduce the overshoot/undershoot voltage in comparison to existing state-of-the-art designs. Moreover, the thesis demonstrates how this structural approach can be extended to efficiently handle extensive distributed loads.

Abstract

General Circulation Models (GCMs) aid in developing climate-resilient policies, preparing for extreme events and implementing governance and disaster management mitigation strategies. Thus, assessing GCMs� climate forecasting capabilities is crucial to establishing the credibility and reliability of climate projections, facilitating well-informed decision-making and climate change readiness. The UK Met Office Hadley Centre Coupled Model, version 3 (HadCM3), is extensively evaluated in this study. HadCM3 is used to simulate global and regional climate patterns and has been crucial to climate change assessments. This thesis explores the spatio-temporal dynamics of temperature in contiguous Peninsular India by identifying regions of high & low variability, with an emphasis on Elevation-Dependent Warming (EDW). The central objective of this study is to improve our understanding of the attributes, patterns, trends, and fundamental factors that affect temperature variations across diverse spatial and temporal dimensions and the HadCM3 model�s ability to accurately represent this phenomenon. Through a 30-year analysis of observed and modelled minimum (Tmin) and maximum temperature (Tmax) data (1991-2020 and 1990-2019, respectively) over the study region, this study explores the relationship between elevation and warming trends. The observed and modelled Tmin and Tmax have been increasing at a rate of +0.13�C/decade and +0.17�C/decade for observed data & +0.43�C/decade and +0.45�C/decade for modelled data, whereas the observed and modelled diurnal temperature range (DTR) increased at +0.03�C /decade & +0.02�C/decade. The lapse rates of observed and modelled minimum and maximum temperatures are found to be positive, but the DTR lapse rate is found to be negative, indicating increasing DTR with elevation. Modelled data shows more pronounced trends (+6.344�C/km for Tmin, +5.954�C/km for Tmax, and -0.39�C/km for DTR) than observed temperature (+2.376�C/km for Tmin, +2.341�C/km for Tmax, and -0.193�C/km for DTR). This discrepancy suggests that the global average lapse rate (+6.5�C/km) underlines the modelled data, resulting in much higher lapse rates in the study region, which implies that the model uses a uniform lapse rate to model temperature and fails to account for the study region�s unique topographic temperature relationships. The study proposes the P-MiSTIC method, a multivariate extension of MiSTIC, to identify zones with spatio-temporal consistency and investigate these inconsistencies. This study uses Tmin and Tmax as P-MiSTIC inputs with weights of -1 and 1. Using MiSTIC, the study finds 12, 9, 8, and 4 spatio-temporally invariant zones for observed and modelled Tmin and Tmax, respectively. Importantly, the elevation-based change rates of variables for these zones (-2.57�C/km and -2.19�C/km for observed Tmin and Tmax, and -7.48�C/km and -6.67�C/km for modelled Tmin and Tmax) closely match the data-derived lapse rates. Using the P-MiSTIC method, 11 and 1 zones are identified for observed and modelled data, respectively, indicating that modelled data is consistent over the Peninsular region. DTR change rates for zones within observed temperature data increase with elevation, indicating elevation-dependent warming in specific study regions. In particular, the Western Himalayan region and the Karakoram region, with the highest elevations, drive the phenomenon in the study region with the Western Himalayan and Karakoram DTRs increasing at 0.19�C/decade and +0.28�C/decade over the study period. The GCM model�s DTR data shows a much slower increase of +0.02�C/decade in the study region. Using the P-MiSTIC method on three decade-wise subsets, only one zone is identified in modelled data across all decades, while observed data identifies 12, 11, and 9 zones for decades 1, 2, and 3. The elevation-based DTR change rates for the observed temperature zones are estimated at -0.03�C/km, -0.028�C/km, and +0.533�C/km for decades 1, 2, and 3. DTR trends change from decreasing to increasing as the study progresses from decades 1 and 2 to decade 3, suggesting elevation-dependent warming in the study region is accelerating. The Western Himalayas and the Karakoram experience considerable variations of DTR. In the Western Himalayas, the DTR decreases at -0.14�C/decade, -0.29�C/decade, and - 0.41�C/decade for decades 1, 2, and 3, while in the Karakoram, it increases at +0.11�C/decade, +0.13�C/decade, and +2.46�C/decade. However, the model�s DTR data shows a uniform increase of +0.02�C/decade across all decades, indicating the model�s uniform rate. By dividing the datasets into DJF, MAM, JJA, and SON seasons, the study examined seasonal spatio-temporal variations. Seasonal observations identified 11, 10, 7, and 8 zones where the diurnal temperature range (DTR) changed with elevation at -0.518�C/km, -0.089�C/km, 0.668�C/km, and 0.453�C/km for the four seasons. Western Himalayan and Karakor

Abstract

The goal of rendering is to produce a photorealistic image of the given 3D scene description. Physically based rendering simulates the physics of the light as it travels and interacts with objects in the scene before finally reaching the camera sensor. Monte Carlo methods have been the go-to approach for physically based rendering. They are general and robust but introduce noise and are computationally expensive. Recent advancements in hardware, algorithms, and denoising techniques have enabled realtime applications of Monte Carlo methods. However, complex scenes still demand high sample counts. In this thesis, we explore and present the utilization of analytic and neural approaches for physically based rendering. Analytic methods offer noise-free renderings but are less general and may introduce bias. In recent years, neural-based approaches have gained traction, offering a balance between generality and computational efficiency. We compare and contrast the traditional Monte Carlo-based methods and emerging analytic and neural network-based methods. We then propose analytic and neural solutions to two challenging cases: direct lighting with many area lights and efficient rendering of glinty appearances on specular normal-mapped surfaces. Direct lighting from many area light sources is challenging due to variance from both choosing an important light and then a point on it. Existing methods weigh the contribution of all lights by estimating their effect on the shading point. We propose to extend one such method by using analytic methods to improve the estimation of the light�s contribution. This enhancement accelerates the convergence of the algorithm, making it more efficient for scenes with many dynamic lights. The second case deals with the challenge of rendering glinty appearances on normal mapped specular surfaces efficiently. Traditional Monte Carlo methods struggle with this task due to the rapidly changing spatial characteristics of microstructures. Our solution introduces a novel method supporting spatially varying roughness based on a neural histogram, offering both memory and compute efficiency. Additionally, full direct illumination integration is computed analytically for all light directions with minimal computational effort, resulting in improved quality compared to previous approaches. Through comprehensive analysis and experimentation, this thesis contributes to the advancement of rendering techniques, shedding light on the trade-offs between different methods and providing insights into their practical applications for achieving photorealistic rendering

Abstract

Artificial intelligence (AI) has infiltrated all fields of science, from high-energy particle physics to biology to computational chemistry. In the last couple of decades, there has been tremendous advancement in machine learning (ML) applications in computational chemistry. Deep learning (DL) has achieved some success in the automation of feature design, physicochemical property prediction, accelerated chemical space search, and the design of new drug-like molecules. Much work is still needed in terms of property prediction of inorganic molecules, along with the search and design of new molecules and material design with desired properties. This research aims to develop machine learning methods for 3D structure generation and molecular geometry optimization. Use of neural network potential (NNPs) can accelerate the process of 3D structure generation and molecular geometry optimization. Various neural network potentials (NNPs) have been reported in the literature to be as fast as force fields and as accurate as DFT. There has been a lack of standard comparative evaluation of these NNPs, which motivated us to do a benchmark study on NNPs. In this benchmark study, we evaluate and compare four NNPs, i.e., ANI, PhysNet, SchNet, and BAND-NN, for their accuracy in energy prediction, transferability to larger molecules, ability to produce accurate PES, and applicability in geometry optimization. In the context of 3D structure generation (Molecules and material design), there are two major components: search algorithm and property predictor. We need a fast and accurate method to predict the energy of the given system to accelerate the search in conformational space. For this, we developed a model known as DART, which predicts the energy of Gallium clusters using a Topological Atomistic Descriptor (TAD). TAD is a very simple and elegant descriptor that tries to encode structural information by dividing the connectivity information using distance cutoffs. We show the DART models ability to predict the energies of Gallium clusters accurately. For the second component, i.e., the search algorithm, we developed an RL-based model, MeGen, to generate 3D low-energy isomers of Gallium clusters, which uses DART as a reward function. Here we showed that MeGen is significantly more efficient than the conventional workflow for generating ground-state geometries as well as low-lying isomers in terms of time and computational resources. Following a similar train of thought, we developed the MolOpt model. This multi-agent RL-based search algorithm can perform molecular geometry optimization (MGO) by searching for the low-energy structure on the potential energy surface. We show that MolOpt trained on ethane and butane can be used to optimize larger alkanes up to octane. We compare our model with other optimizers and show that MolOpt outperforms the MDMin optimizer and performs similarly to the FIRE optimizer. We further developed an improved version of MolOpt known as MolOpt2. We have made algorithmic changes in MolOpt2, and MolOpt2 is trained on a diverse set of molecules. Hence, due to algorithmic changes, our new model (MolOpt2) can perform MGO on molecules containing elements CHNO and having a size of up to nine heavy atoms. Similar to MolOpt, we compare MolOpt2 with other optimizers and show that MolOpt2 outperforms the MolOpt, MDMin optimizer and performs similarly to the FIRE optimizer.

Abstract

Think of a situation where you put yourself in the shoes of a visually impaired person who wants to buy an item from a store or a person who is sitting in their house and watching the news on the television and wants to know about the content of the news being broadcast. Motivated by many more such situations where creating systems capable of understanding and reasoning over textual content in the videos, in this thesis, we tackle the novel problem of text-based video question answering. Vision and Language are broadly regarded as cornerstones of intelligence. Though each of these has different aims – language has the purpose of communication, and transmission of information, and vision has the purpose of constructing mental representations of the scene around us to navigate and interact with objects. When we study both of these fields jointly, it can result in applications, tasks, and methods that, when combined go beyond the scope compared to when they are used individually. This inter-dependency is being studied as a newly emerging area of a study named “multi-modal understanding”. Many tasks such as image captioning, visual question answering, video question answering, text-video retrieval, and more fall under the category of multi-modal understanding and reasoning tasks. To have a system that can reason over both text-based information and temporal-based information, we propose a new task. The first portion of this thesis focuses on the formulation of the text-based VideoQA task, by first analyzing the current datasets and works and thereby arriving at the need for text-based VideoQA. To this end, we propose the NewsVideoQA dataset where the question-answer pairs are framed on the text present in the news videos. As this is a new task proposed, we experiment with existing methods such as text-only models, single-image scene text-based models, and video question-answering models. As these baseline methods were not originally designed for the task of video question-answering using text in the videos, the need for a video question-answering model that can take the text in the videos into account to obtain answers became the need. To this end, we repurpose the existing VideoQA model to incorporate OCR tokens namely – OCR-aware SINGULARITY, a video question-answering framework that learns joint representations of videos and OCR tokens at the pretraining stage and also uses the OCR tokens at the finetuning stage. In this second portion of the thesis, we look into the M4-ViteVQA dataset which aims to solve the same task of text-based video question-answering but the videos belong to multiple categories such as shopping, traveling, vlogging, gaming, and so on. We perform a data exploratory analysis where we analyze both NewsVideoQA and M4-ViteVQA on several aspects that look for limitations in these datasets. Through the data exploratory experiment, we show that most of the questions in both datasets have questions that can be answered just by reading the text present in the videos. We also observe that most of the questions can be answered using a single to few frames in the videos. We perform an exhaustive analysis on a text-only model: BERT-QA which obtains comparable results to the multimodal methods. We also perform cross-domain experiments to check if training followed by finetuning on two different categories of videos helps the target dataset. In the end, we also provide some insights into creating a dataset and how certain types of annotations can help the community come up with better datasets in the future. We hope this work motivates future research on text-based video question-answering in multiple video categories. Furthermore, the pretraining strategies and combined representation learning from these videos and the multiple modalities that videos provide us will help create scalable systems and drive future research towards better datasets and creative solutions.

Abstract

Wikipedia is one of the primary sources of encyclopedic content online. It is one of the most widely read websites. It is also regarded as a quality data source in many machine-learning pipelines. To maintain the high quality of its articles, Wikipedia has three core content policies, one of which is “Neutral Point of View (NPOV)”. This policy is a set of principles, including “avoiding stating opinions as facts” and “preferring nonjudgmental language.” Whenever we refer to “bias”, we refer to it within these guidelines. This work studies how to enhance the quality of the Indian language Wikipedia articles. We looked at existing work on dataset curation from English Wikipedia for bias detection and tried replicating that for Indian languages. We discuss the hurdles faced in this process and discuss translation (along with various quality checks to reduce noise) as a viable alternative. Much of this thesis is dedicated to automatically detecting whether a sentence can be called biased and trying to remove the bias if so. Bias detection is challenging because certain words lead to bias if written in some contexts while not in others. For bias detection in Indian languages, we perform binary classification using MuRIL, InfoXLM and mDeBERTa in zero-shot, monolingual and multilingual settings. For human evaluation, we note how this is a subjective task and disagreement among annotators is expected. Thus, we also experiment with different settings like loss functions specific for subjective tasks and include anonymized annotator-specific information to help us understand the level of disagreement. For bias mitigation, we perform style transfer using IndicBART, mT0 and mT5. These models provide strong baseline results for the novel multilingual tasks. We study how different text generation metrics may or may not be able to capture the quality of debiasing and how to evaluate our models best. Reinforcement learning offers a way to fix the problems observed in the debiased results of the style transfer module. Also, it helps us combine the classification and style transfer modules. We formulate three reward functions specific to our debiasing task and study the results of training fully/partially with these rewards compared to vanilla mT5. Yet another way of improving the quality of the Indian language Wikipedias is to verify the accuracy and reliability of the information presented. All material in Wikipedia must be attributable to a reliable, published source. Thus, we try to identify if the information in a sentence is factually correct or needs a citation. In contrast to previous work, we do this at the fact level instead of the sentence level for more accurate results. Thus, these measures will enhance the quality of the Indian language Wikipedia articles and increase its credibility as the largest source of free, fair, and accurate information.

Abstract

Consider an autonomous agent capable of obeying the following instruction �Go and clean the coffee spilled on the dining table�. To perfectly execute this instruction, the agent needs to have a precise understanding of its dynamic 3D environment. The final task can be broken down into the following sub-tasks; 3D grounding, 3D semantic understanding, and 3D motion planning. To excel in all these tasks, 3D representation learning plays an important role. Motivated to contribute towards creating systems that can perceive and act in real 3D words, in this thesis, we propose two novel methods for 3D representation learning. The first portion of this thesis focuses on the tasks of unsupervised domain adaptation (UDA) for 3D point clouds. The point cloud data acquisition procedures manifest themselves as significant domain discrepancies and geometric variations among both similar and dissimilar classes. The standard domain adaptation methods developed for images do not directly translate to point cloud data because of their complex geometric nature. Existing works mainly focused on designing a self-supervised task to improve adaptation performance. We propose a new UDA architecture for point cloud classification that benefits from multimodal contrastive learning to get better class separation in both domains individually. Further, the use of optimal transport aims at learning source and target data distributions jointly to reduce the cross-domain shift and provide a better alignment. We conduct a comprehensive empirical study on PointDA 10 and GraspNetPC-10 and show that our method achieves state-of-the-art performance on GraspNetPC-10 (with ? 4-12% margin) and best average performance on PointDA-10. Our ablation studies and decision boundary analysis also validate the significance of our contrastive learning module and OT alignment. In the second portion of the thesis, we explore learning Wasserstein embeddings for point clouds towards multiple downstream tasks. As learning embeddings of any data largely depends on the ability of the target space, we propose to embed point clouds as discrete probability distributions in Wasserstein space. We build a contrastive learning setup to learn Wasserstein embeddings that can be used as a pre-training method with or without supervision towards any downstream task. We show that the features captured by Wasserstein embeddings are better in preserving the point cloud geometry, including both global and local information, thus resulting in improved quality embeddings. We perform exhaustive experiments and demonstrate the effectiveness of our method for point cloud classification, transfer learning, segmentation, and interpolation tasks over multiple datasets, including synthetic and real-world objects. We also compare against recent methods that use Wasserstein space and show that our method outperforms them in all downstream tasks. Additionally, our study reveals a promising interpretation of capturing critical points of point clouds that makes our proposed method self-explainable. We hope this work motivates future research in utilizing optimal transport for understanding the real 3D world. We also hope that the self-supervised approaches proposed in this thesis will act as a step towards in this direction.

Abstract

The convergence of Deep Reinforcement Learning (DRL) and cognitive neuroscience has opened avenues for probing neural representations within the brain. This thesis delves into the synergy between decision-making and representation by harnessing DRL’s capabilities. Through a focus on dot motion perceptual decision-making task within a high-dimensional framework akin to psychological experiments, we acquire novel insights into how intricate neural networks tackle complex tasks. The resultant end-to-end model not only elucidates network strategies, but also offers a backdrop for understanding analogous brain processes. Investigation reveals intriguing resemblances between the network’s behavior and neural mechanisms observed in the middle temporal visual area (MT) of the brain, known for encoding direction and motion strength. Remarkably, direction selectivity emerges solely through reward-driven training, while graded firing patterns encode motion strength. A testable hypothesis arises, suggesting coherenceselective neurons within the MT population. While conventional Reinforcement Learning (RL) presupposes uniform data distributions, real-world scenarios like autonomous driving and natural environments display Zipfian distributions, characterized by frequent occurrences amidst rare events. Inspired by complementary learning systems theory, this work introduces an architecture tailored for learning from Zipfian distributions. Unsupervised discovery of long tail states and their retention in episodic memory, coupled with recurrent activations, forms the core of the proposed architecture. Retrieval from episodic memory via similarity-based search, reinforced with weighted importance, amplifies performance across diverse Zipfian tasks. Our proposed architecture achieves higher accuracy than the IMPALA algorithm across all three tasks and evaluation metrics (Zipfian, Uniform, and Rare Accuracy). This thesis advances the nexus of decision-making and neural representation through DRL, while proposing an innovative architecture capable of navigating the intricacies of Zipfian data distributions. It not only augments our grasp of cognitive building blocks but also bears implications for resolving challenges across real-world domains. This research in one small step that bridges the gap between computational models and cognitive processes, opening new vistas for understanding and application.