Abstract

Typically, anatomy is taught through dissection, 2D images, presentations, and cross-sections. While these methods are convenient, they are non-interactive and fail to capture the spatial relationships and functional aspects of anatomy. New visualization technologies such as virtual reality and 3D can compensate for the impediments and provide better understanding while captivating the students. With recent advances in the industry, the methods to provide a 3D experience are economical. In this thesis, we introduce a low-cost 3D-interactive anatomy system designed for an audience of typical medical college students in an anatomy class. The setup used to achieve 3D visualization is Dual projector polarization. While there are other ways to achieve 3D visualization, like alternate frame sequencing and virtual reality, this technique can target a large audience and requires minimum accessories for the setup enabling this to be a low-cost solution for an immersive 3D experience. The 3D interactive Neuroanatomy solution is an end-to-end framework capable of designing anatomy lessons and visualizing the 3D stereoscopic projection of those anatomy lessons. To ensure superior comprehension of students, we incorporate each teacher’s unique teaching approach while developing anatomy lessons by providing the ability to create their own lessons. We have created anatomy lessons based on the human brain which is a vital organ and has a complex anatomy. Our aim is to help medical students to understand the complexity of organ systems from not just an anatomical perspective but also a radiological perspective. We use annotations on clinical case data such as MRI, MRA, etc., to create 3D models for anatomy visualization incorporating clinical information and illustrating real cases. Annotations for structures of interest are done using manual, automatic, and semi-automatic segmentation methods. Manual delineation of the structure boundaries is very tedious and time-consuming. Automatic segmentation is quick and convenient. However, manual annotations were done for the 3D anatomy viewer for small and complex structures due to substandard automatic segmentation. There is a need to improve automatic segmentation performance for those structures. While segmentation is an essential step in 3D modeling, it plays a critical role in many neurological disease diagnoses as well, which are associated with degradation in the sub-cortical region. Therefore accurate algorithms are needed for sub-cortical structure segmentation. Variance in the size of structures is significant, which introduces a performance bias towards larger structures in many deep learning approaches. In this part of the thesis, we aim to remove size bias in sub-cortical structure segmentation. The proposed method addresses this problem with a pre-training step used to learn tissue characteristics, an ROI extraction step that aids in focusing on the local context, and using structure ROIs elevates the influence of smaller structures in the network.

Abstract

FPGAs are increasingly significant for deploying convolutional neural network (CNN) inference models because of performance demands and power constraints in embedded and data centre applications. The compute intensity of these models makes prototyping highly complex and time-consuming with traditional RTL approaches. The release of new generation high-level synthesis tools (HLS), such as Intel FPGA SDK for OpenCL, and Xilinx’s VITIS Unified Software Development platform, have significantly reduced the time and complexity of prototyping complex designs on FPGA. This work involves building custom FPGA accelerators for image recognition systems using OpenCL-HLS. Object detection and classification are vital steps in building image recognition systems. The first part of the work concerns building an FPGA accelerator for the Traffic Sign Classification problem, a vital step in building traffic sign recognition (TSR) systems that employ vehicle-mounted cameras that identify traffic signs while driving on the road. However, the CNNs for the classification still need the ability to be spatially invariant to the input data. Spatial Transformers are learnable modules that, upon integration with CNN, would allow the spatial manipulation of data within the network, making it invariant to affine transformations. Generic Matrix multiply (GEMM) methods that express convolution as matrix multiplication are widely used in deep-learning frameworks like Caffe, Theano, and Torch with GPU support. im2row is one of the commonly used GEMM methods. In this work, we built a GEMM-based accelerator for a CNN with a Spatial transformer module. We proposed the channel adaptive im2row method, with a lesser on-chip memory footprint than im2row. The system attains a latency of 202 ms (5˜ fps), running at 202 MHz on Intel Arria10 GX FPGA, and attains a speedup of (> 5 X) compared to the CPU. The performance is not state-of-the-art and calls for more FPGA-specific optimizations. Further, from the learnings, we designed a Systolic Array accelerator with a novel load pattern for accelerating the widely-used object detector YOLOv3-tiny optimized explicitly for embedded applications. We build the accelerator for multiple precisions ( FIXED8, FIXED16, FLOAT32 ) of YOLOv3-tiny. The architecture uses a homogenous systolic array architecture with a synchronized pipeline adder tree for convolution, allowing it to be scalable for multiple variants of YOLO with a change in the host driver. It is a deeply pipelined architecture that also exploits three-dimensional spatial parallelism. We evaluated the design on Terasic DE5a-Net-DDR4. The Fixed point (FP-8, FP-16) implementations attain a throughput of 57 GOPs/s (> 23 %) and 46.16 GOPs/s (> 340 %) running at 234 MHz and 227 MHz. We synthesized the first FLOAT32 implementation attaining 11.22 GFLOPs/s, running at 172 MHz

Abstract

Deep Neural Networks (DNNs) have shown remarkable performance in a broad range of computer vision tasks, including in the medical domain. With the advent of DNNs, the medical community has witnessed significant developments in segmentation, classification, and detection. But this success comes with a cost of heavy reliance on the abundance of data. Medical data, however, is often highly limited in volume and quality due to sparsity of patient contact, variability in medical care, and privacy concerns. Hence, to train large networks we seek data from different sources. In such a scenario, it is of interest to design a model that learns continuously and adapts to datasets or tasks as and when they are available. However, one of the important steps to achieve such a never-ending learning process is to overcome Catastrophic Forgetting (CF) of previously seen data or tasks. CF refers to the significant degradation in performance on the old task/dataset. To avoid confusion, we call a training regime Continual Learning (CL) when CAD systems have to handle a sequence of datasets collected over time from different sites with different imaging parameters/populations. Similarly, Incremental Learning (IL) is when CAD systems have to learn new classes as and when new annotations are made available. The work described in this thesis address core aspects of both CL & IL and has been compared against the state-of-the-art methods. In this thesis, we assume that access to the data belonging to previously trained datasets or tasks is not available which makes both CL and IL processes even more challenging. We start with developing a CL system that learns sequentially on different datasets and handles CF using the Uncertainty mechanism. The system consists of an ensemble of models which are trained or finetuned on each dataset and considers the prediction from the model which has the least uncertainty. We then investigate a new way to tackle CF in CL by manifold learning, inspired by the defense mechanisms against adversarial attacks. Our method uses a ‘Reformer’ which is essentially a denoising autoencoder that ‘reforms’ or brings the data from all the datasets together towards a common manifold. These reformed samples are then passed to the network to learn the desired task. Towards IL, we propose a novel approach that ensures that a model remembers the causal factor behind the decisions on the old classes, while incrementally learning new classes. We introduce a common auxiliary task during the course of incremental training, whose hidden representations are shared across all the classification heads. All the experiments for both CL and IL are conducted on multiple datasets and have shown significant performance over the state-of-the-art methods

Abstract

The conformational dynamics of proteins is essential for biological functions. The characterization of binding-induced changes in protein conformational dynamics is of paramount importance to understand the fundamental processes of life. Molecular modelling and molecular dynamics simulations have emerged as useful computational tools to probe atomistic details of molecular recognition and binding-induced dynamical responses of proteins. The interactions of proteins with DNA and small drug molecules play crucial roles in regulating various cellular processes. The present thesis reports results of investigation of the functional roles of protein conformational dynamics in the following two important biological systems: (a) a DNA repair protein complexed with a mistmatched DNA and (b) an inhibitor drug-bound human c-Src kinase. In the former system, our primary objective is to understand the molecular mechanism of how a specific DNA repair protein (RAD4/XPC) binds to a damaged DNA, recognizes lesions and repair them efficiently using molecular dynamics and enhanced sampling simulations. In the latter system, we have investigated drug-induced conformational transitions in human c-Src kinase to understand its activation mechanism using molecular dynamics simulations and nudged elastic band method. DNA damage caused by ultraviolet (UV) radiation can lead to a range of genetic skin conditions and cancers. The protein Rad4/XPC recognizes and repairs these types of DNA damage with high accuracy to safeguard the integrity of the genome. However, the precise mechanism of how Rad4/XPC recognizes DNA damage in the crowded cellular milieu is not yet fully understood. The first part of this thesis investigates the mechanism, energetics, dynamics, and molecular basis of Rad4/XPC’s recognition with base pair mismatched DNA lesions. In particular, the mechanism of association of RAD4 with the mismatched DNA is investigated using molecular dynamics and umbrella sampling simulations. The Rad4-DNA association free energy surfaces are determined for three Rad4-DNA complexes (CCC/CCC, TTT/TTT, and TAT/TAT) and the sequence-dependent specificity of Rad4 for mismatched DNA is explored. The results reveal that Rad4 exhibits higher specificity for CCC/CCC mismatched DNA compared to the other two mismatches. The key molecular interactions contributing to Rad4- DNA association and specificity are characterized. Protein kinases act as biological switches that toggle between ON (active) and OFF (inactive) states in response to specific cellular signals to regulate various cellular processes including growth, cell proliferation and differentiation. Since protein kinases are implicated in tumorigenesis and malignant cell proliferation, they have become favourable targets for cancer therapy. The second part of the thesis focuses on drug-induced conformational transition of c-Src kinase between active (drug-bound) and inactive (drug-free) states using the minimum energy path analysis and molecular dynamics simulation. The crucial role of hydration of the binding pocket in the association/dissociation of drug from the kinase is explored. The coupling between the entry and exit of the drug, influx of water molecules into the binding pocket and critical interactions of key kinase residues near the binding pocket appears to influence the conformational transition between the active and inactive states of the kinase. We believe that the molecular mechanisms elucidated will be useful for rational design of novel therapeutics against cancer and other kinase-related diseases.

Abstract

Drought is a natural hazard with a significant impact on the economy, agriculture, and environment. It is defined as a significant decrease in water availability in all its forms. Droughts are estimated using drought indices. Drought indices are numerical measurements that describe the severity of drought by combining data from one or more variables (indicators), such as precipitation and evapotranspiration (ET), into a single number. Standardized Precipitation Evapotranspiration Index (SPEI) and Standardized precipitation Actual Evapotranspiration Index (SPAEI) are the drought indices used to estimate drought index, which take both precipitation, potential and actual evapotranspiration into account. Since ET is a critical factor in estimating drought, well-grounded ET estimations are required. The major forms of ET are Potential evapotranspiration (PET) and Actual evapotranspiration (AET). PET is described as the loss of water from a significant area that is equally covered with short, green crops that are actively growing. PET is regarded as the maximum amount of evapotranspiration. AET is defined as the total amount of water used in evaporation and transpiration by a crop during the entire growing season. Various methods have been developed to estimate PET and AET depending upon the availability of hydro meteorological variables. Various empirical based methods and hydrological model based simulations of PET and AET have been developed. Advancement of data-driven algorithms also have been extensively developed to estimate ET. In this context, many studies used empirical based estimates of ET for the calculation of drought indices. None of these studies analyzed the sensitivity of various ET methods to assess the drought characteristics. Thus, the present study aimed to include various PET and AET methods in the drought characterization. Various empirical methods, such as Penman-Monteith, Hargreaves, Turc, and Priestley-Taylor and data-driven method of Artificial Neural Network (ANN) has been used to estimate PET. The input variables used to estimate these methods are temperature, wind speed, solar radiation, and relative humidity which are obtained over the period of 1965 to 2015 for Hyderabad station. ANN model was trained and tested with climate variables as input variables and various empirical models as reference models to predict the best PET method. Penman-Monteith, Hargreaves, and Turc method performed better with ANN model estimates. Later, to study the impact of various empirical based PET estimates on drought estimation, SPEI is calculated using various PET estimates at different time scales. All other methods are compared with the Penman-Monteith method, which is considered the standard method because it considers the main meteorological factors. Hargreaves and Turc methods performed better with the standard method and these methods can be useful in estimating drought when minimum data is available. To assess the drought events accurately by various drought indices it is necessary to predict the hydro-meteorological variables (PET and AET) precisely. There are several challenges in estimating AET and PET at the fine spatial resolution. There are various empirical models (Budyko, Penman- Monteith, Hargreaves, and Turc) for estimating AET and PET. Still, these empirical methods does not account for the catchment characteristics, which may underestimates the actual amount of hydrological variables. Further, satellite-based remote sensing data are accessible for extracting evapotranspiration (ET) values. It provides global coverage and continuous observations of land surface variables affecting ET. Another conceptual based approach to estimate PET and AET at catchment scale is hydrological model such as Soil and Water Assessment Tool (SWAT). The present study aimed to include various approaches of empirical (Budyko, Penman-Monteith, Hargreaves, and Turc), modeled (SWAT), and remote sensing in the drought characterization using SPEI and SPAEI. Remote sensing PET and AET are considered as standard methods to compare both empirical and modeled PET and AET estimates. The present methodology was tested on a dry-sub-humid river catchment of India, the Tungabhadra River catchment. It is recommended to use PET instead of AET when estimating drought indices as SPEI values performed relatively better than SPAEI. In the present study it is observed that although PET and AET estimates vary with different models, drought indices SPEI and SPAEI are not differing much at annual scales. Hargreaves and Penman-Monteith performed better results compared to remote sensing method in SPEI calculations. And for SPAEI, Budyko and Turc performed better results. Thus, the present study concludes that empirical models correlated better with the remote sensing data. The study will be prominent for ungauged river basins, where detailed hydrological data is limited and difficult

Abstract

To perform network forensic analysis on an incident in software defined networking (SDN), logs are of the utmost importance. Without any logs, an investigator will not be able to complete and justify the forensic analysis. With the advancement of network and communication technologies, the volume of logs that needs to be collected and retained is growing exponentially big. The network administrators face the problems in maintaining such huge logs for longer period of time for forensic analysis. Network Providers rely on purging data based on a certain stipulated duration expected by the local rules for forensics evidence. SDN providers have limitations in managing such big data since the expense involved is commensurate with the duration of data retention. Here, in this work we propose a novel idea to reduce and summarize large forensic data sets for faster querying as well as reducing its space complexity by using bloom filters. Through this work we aim to propose a system which can deliver more optimized forensic data availability in SDN platform compared to existing systems.

Abstract

In recent years, sentiment analysis has gained popularity as it is essential to moderate and analyse the information across the Internet. It has various applications like opinion mining, social media monitoring, and market research. Aspect Based Sentiment Analysis (ABSA) is an area of sentiment analysis which deals with sentiment at a finer level. ABSA classifies sentiment with respect to each aspect to gain greater insights into the sentiment expressed. Significant contributions have been made in ABSA, but this progress is limited only to a few languages with adequate resources. Telugu lags behind in this area of research despite being one of the most spoken languages in India and an enormous amount of data being created each day. In this thesis, we create a reliable resource for aspect based sentiment analysis in Telugu. The data is annotated for three tasks namely Aspect Term Extraction, Aspect Polarity Classification and Aspect Categorisation. Further, we develop baselines for the tasks using deep learning methods demonstrating the reliability and usefulness of the resource. In addition to this, experimentation has been done with transformer models, as they have led to pivotal changes in the field of NLP. Specifically, the setting of monolingual and multilingual models has been chosen to compare the effective contribution of the dataset created for ABSA in Telugu.

Abstract

Although the concept of a knowledge graph has been discussed since at least 1972 [93] it wasn’t until Google [96] unveiled their version in 2012 that it truly took off. Since then, several companies have started working on their own knowledge graphs, including Google, Amazon [54], eBay [82], Twitter, IBM [21], LinkedIn [37], Microsoft [95], and Uber. There has been a rise in the number of academic publications devoted to the topic of knowledge graphs in recent years, reflecting the growing interest in this idea. There are several books [79, 85, 52] and papers [84, 24] that detail knowledge graphs, as well as unique techniques to generating and analysing knowledge graphs and assessments of various knowledge graph aspects [103]. Apart from these enterprise Knowledge Graphs which are private and not accessible to public, there are a number of a number of public knowledge graphs published where the content is accessible for users of the web. The publicly available Knowledge Graphs are called open KGs. DBpedia [58], YAGO [40], Freebase [101], Wikidata [8] are the top few examples of such open KGs which are multi-domain and are built using the data from Wikipedia. Automatic extraction of information from text and its transformation into a structured format is an important goal in both Semantic Web Research and computational linguistics. Knowledge Graphs (KG) serve as an intuitive way to provide structure to unstructured text. A fact in a KG is expressed in the form of a triple which captures entities and their interrelationships (predicates). Multiple triples extracted from text can be semantically identical but they may have a vocabulary gap which could lead to an explosion in the number of redundant triples. Hence, to get rid of this vocabulary gap, there is a need to map triples to a homogeneous namespace. In this work, we present an end-to-end KG construction system, which identifies and extracts entities and relationships from text and maps them to the homogenous DBpedia namespace. For Predicate Mapping, we propose a Deep Learning architecture to model semantic similarity. This mapping step is computation heavy, owing to the large number of triples in DBpedia. We identify and prune unnecessary comparisons to make this step scalable. Our experiments show that the proposed approach is able to construct a richer KG at a significantly lower computation cost with respect to previous work. Over recent years, document similarity has grown to become the foundation of various natural language processing activities, which are crucial to information retrieval, automatic question answering, machine translation, dialogue systems, and document matching. For document or topic similarity, focusing on the semantics within the text has been the most common and pursued direction of effort Our novel KG-based similarity classifier works on the limitations of previous approaches and also provides reasoning behind the classification. Our results show that we’re able to score similarity between Wikipedia documents accurately. Furthermore, the accuracy of our approach is able to withstand against noise and paraphrasing. We also see that our classifier can be used for category outlier detection in DBpedia. In this thesis, we will focus on Knowledge Graph construction from unstructured text and document similarity. Our knowledge graph construction approach performs better > 25% better than Cosine & Rule-based approaches, and is also computationally cheaper than state-of-the-art T2KG [53] by a magnitude of atleast 106 . We use these constructed knowledge graphs to classify similarity between two documents, which is used to detect category outliers in DBpedia and find highly similar documents.

Abstract

Urdu writing system is derived from the Persio-Arabic writing systems and thus it has adopted similar orthographical and morphological characteristics as that of Persio-Arabic languages. The first and foremost task for most of the NLP applications is Word Segmentation which involves identifying the bounding boundaries of words in written text. It is quite crucial to accurately identify the boundaries of each word in written text because all the downstream tasks in NLP are dependent on it, thus making Word Segmentation fundamentally important. Urdu adopts a continuous writing style which does not have an explicit and clear marker for word boundary. Furthermore, the inherent non-joining attributes of certain characters in Urdu create spaces within a word while writing in digital format. Thus, Urdu not only has space omission but also space insertion issues which make the word segmentation task challenging. We have studied and categorized the various issues that are observed with respect to the inconsistent usage of space character in Urdu script along with the orthographic and morphological reason behind it. Another challenge in computational processing of Urdu is the lack of benchmark resources and corpora for Word boundary identification. Leveraging the learning from the orthographic study of Urdu writing system, we have built a benchmark corpus for Urdu Word Segmentation, with an exercise of manual annotation, using white space as word boundary and Zero-Width-Non-Joiner (ZWNJ) character as sub-word boundary. A Conditional Random Field based sequence modeler was then used to train a character-level label prediction of a sequence of Urdu characters. Our model achieved state-of-the-art results with an F1 score of 0.98 for word boundary identification. Furthermore, we have applied our word segmentation model on studying the sociological phenomena of Diglossia in Urdu

Abstract

Gas sensors find widespread applications from healthcare to industrial safety. The detection of volatile organic compounds like acetone and ethanol in their gaseous state is very important, especially in the biomedical field. SMO-based gas sensors, which have simple architecture and several favorable features, require additional treatments like noble metal loading onto the SMOs like ZnO surface for superior performance. Experimental trial-and-error approach to performance improvement has been the prevalent practice due to the lack of knowledge of the electronic and chemical properties of such systems. Computational studies can play significant roles as predictive tools in this field. In this work, by using DFT study, we explore how Cu, Ag, Au, or Pd nanocluster loading onto ZnO cluster affects acetone vs. ethanol sensing performance of ZnO semiconductor. DFT simulations have been used obtain the chemical and electronic properties of various nanoclusters systems, which have been associated with sensor performances. First, we have studied the binding of acetone and ethanol to pristine (ZnO)2 cluster which explains how ZnO show better sensor response to acetone than ethanol. Next, we have explored the metal nanocluster M6 (where M = Cu, Ag, Au, and Pd) loading onto ZnO. Since the actual working motifs in the nanostructured sensor materials could be a small cluster of few atoms, we have taken M6 nanocluster as a model for the metal nanoparticle. Finally, we have investigated acetone vs. ethanol sensing performance of the metal nanocluster loaded ZnO systems. For each system, we have calculated various parameters like the binding energy, HOMO-LUMO energy and bandgap, Mulliken charge distributions, bond distance, etc. The study shows that the electronic structure effects of M6/ZnO nanoclusters can be used to understand the molecular level mechanism of sensor activation by metal nanoclusters and explain their acetone vs. ethanol sensing performance. A better understanding of the roles of various components of the sensor materials will help in better selection of sensor materials and develop efficient, low cost and portable sensors.