Dissertations - M Tech (CS)

Permanent URI for this collectionhttp://164.52.219.250:4000/handle/10263/2147

These Dissertations were submitted in partial fulfilment of the requirements for the award of M TECH (Computer Science) Degree of Indian Statistical Institute

Browse

Start date: 2010End date: 2019

Search Results

Now showing 1 - 10 of 190
  • No Thumbnail Available
    Item
    Adversarial Attack on Neural Machine Translation System
    (Indian Statistical Institute, Kolkata, 2019-06) Abijith, K P
    Nowadays Deep Neural Network based solutions are deployed to solve numerous tasks. Thus, it has become absolutely important to study the robustness of these systems. Machine Translation is one of the popular applications of Deep Neural Networks. This thesis studies the robustness of Neural Machine Translation systems by generating adversarial examples with the objective to fool the model. Whenever there is a change in the source, i.e. when a word in the input sentence is replaced by an unrelated word, the translation system is supposed to re ect the changes while doing translation. These unwanted invariance learned by the model is undesirable. With intention to exploit this undesirable property learned by a Neural Machine Translation system we design an attack called: Invariance-based targeted attack. This attack introduces multiple changes(replacement of words) to the original input sentence, keeping the translation unchanged. In-order to facilitate the explanation of the design of the attack we introduce two methods: (i) Min-Grad method: To identify the position where a replacement of the word makes the least change in the translation, and (ii) Soft-Attn method: To search for a new word to replace, given a list of choices. The initial part of the report explain the preliminary explorations we did in-order to get some insights on how to do the problem formulation. These experiments are run on LSTM based models with single replacement policy. Using the learning from the rst part we extend the experiments to Transformer and BLSTM based models, which are considered as the state-of-the-art systems for machine translation.
  • No Thumbnail Available
    Item
    Discriminative Dictionary Learning by Exploiting Inter-Class Similarity for HEp-2 Cell Classi cation
    (Indian Statistical Institute,Kolkata, 2019-07) Panda, Aditya
    In this literature we present an algorithm for automatic classi cation of IIF images of HEp-2 cells into relevant classes. Our algorithm is majorly based on the \Dictionary Learning" algorithm and we have rede ned it's objective function to suit our purpose. The major di culty in HEp-2 cell image classi cation lies in it's low inter-class variability and substantial intra-class variations. To address these issues, we have modi ed the objective function of \Dictionary Learning" to learn inter-class features. Moreover, we used a local feature extractor based pre-processing stage and also a \spatial decomposition" classi er set-up for better classifying test images. We evaluated our algorithm on three most widely accepted bamechmark data-sets for HEp-2 cell classi cation, ICPR 2012, ICIP 2013 and SNP data-sets. Proposed algorithm has achieved superior results than other popular dictionary learning algorithms for HEp-2 cell classi cation. Moreover, when comparing with other algorithms for HEp-2 cell classi cation, including the winners of ICPR 2012, ICIP 2013 and SNP data-set, we show that proposed algorithm reports very competitive result. Though our proposed algorithm is designed to be application speci c to HEp-2 cell, still we evaluated its performance on another popular benchmark data-set, \Diabetic Retinopathy" data-set. Our algorithm provided higher accuarcy than other state-ofthe- art algorithms on that data-set too.
  • No Thumbnail Available
    Item
    BlockV: A Blockchain Enabled Peer-Peer Ride Sharing Service
    (Indian Statistical Institute,Kolkata, 2019-07) Pal, Panchalika
    Today's ride sharing is a centralized trust based system where users trust the service providers for the ride set up, tracking, cancellation, fare calculation etc. Any malicious activity in the centralized server based system or a malicious driver or a malicious rider destroys the fairness involved in the ride and causes inconvenience to the parties. After the completion of the ride, the drivers are rated by the riders. There are possibilities that, a malicious rider can claim the refund with a fake complain and give the driver poor rating intentionally or a malicious driver follows a longer path unnecessarily and charges the rider more. Current system is not capable of deciding the correctness of the objections raised by either parties regarding the ride and provides a biased outcome of each objections as per the centralized company's marketing strategies. In this context, we present BlockV, a blockchain enabled anonymous permissionless solution to ensure end to end fairness of the ride. The creation, completion, dissatisfaction or abortion of any ride will be written in the blockchain ledger, hence will be available to all participants in the peer to peer network. This simultaneously ensures the fairness in maintenance of the inbuilt reputation system. We have implemented a prototype in Ethereum private network and KOVAN test network and analyzed the security.
  • No Thumbnail Available
    Item
    Recognition of Strokes in Tennis Videos Using Deep Learning
    (Indian Statistical Institute,Kolkata, 2019-07) Singanporia, Kushal
    Prior introduction of neural nets to domain of computer vision, action recognition requires specific domain knowledge. Still domain knowledge is useful in action recognition but with availability of huge data and neural nets, data-driven feature learning methods have emerged as an alternative. Recent trends in action recognition uses LSTM and its various modifications, as LSTM have memory retaining capability which other architectures lake. In this work we performed action recognition on different tennis strokes. Our work relay on architecture proposed By Husain, Dellen, and Torras, 2016. Architecture is comprised of various modified VGG-nets connected in parallel. As it doesn’t include LSTM, which makes it different than other works.
  • No Thumbnail Available
    Item
    Classi cation of Micro-Blog Texts
    (Indian Statistical Institute,Kolkata, 2019-07) Sen, Bihan
    Classi cation of micro-blog texts is a very common task for sentiment analysis, user opinion mining, product review analysis, crisis managements, identifying ofensive and hate speech propagation across social media, restricting unnecessary expansion of fake news and rumors etc. In this dissertation, we consider two problems from this domain: (i) classi cation of tweets during crisis scenarios like natural disasters, terrorist attacks etc and (ii) identifying o ensive tweets. We tried both statistical and deep learning approaches. Datasets from the TREC-IS 2018 and 2019 tasks, and OLID from O enseEval workshop were used for our experiments. The rst task is formulated as a multi-label classi cation task, while the second is a binary classi cation problem. Our results suggest that preprocessing of social media text is very crucial for classi cation. We also conclude that Deep Learning approaches do not always outperform traditional learning. We also took part as an active participant in the TREC-IS 2019A task. Out of all 34 submissions from across the world, one of our submissions achieved the highest macro-averaged F-1 score on this task (0.1969) and outperformed the second highest score (0.1556) by a substantial margin.
  • No Thumbnail Available
    Item
    On Testing of Samplers and Related Problems
    (Indian Statistical Institute,Kolkata, 2019-07) Chakraborty, Shayak
    The last decade has seen an unprecedented adoption of Arti cial Intelligence and Machine Learning Algorithms in various engineering applications such as medical imaging, autonomous vehicles, web services, etc. Probabilistic Reasoning lies at the heart of such algorithms. Probabilistic reasoning techniques rely highly on sampling techniques and thus samplers that are able to generate quality samples from a discrete distribution uniformly at random, are in heavy demand. This has created a need for e cient samplers that come with some correctness and performance guarantees, and indeed, a number of proposals to this e ect have been developed in literature. This has also created a need to nd an e cient and scalable way for testing the reliability of a sampler with respect to the quality of the samples it generates. Unlike in the eld of program testing, in which nding a single trace of execution is su cient to prove the existence of a bug, the number of samples needed for sampler testing is not one, in fact, most of the testing mechanisms existing in literature, use exponential or sub-exponential number of samples to test for uniformity of a given sampler. The objective of this work is to propose techniques that can cut down on this complexity for certain classes of samplers. To overcome the high sample complexity for property testing of distributions, a framework for conditional sampling was proposed for checking properties of distributions, with polynomial complexity in the number of dimensions of the sample space. A recent framework named Barbarik that works on the guidelines of conditional sampling has suggested a way to test uniformity of samplers sampling from Boolean domains, using a constant number of samples. Barbarik has been proposed in the context of uniformity testing of samplers which claim to provide uniform witnesses for arbitrary Boolean formulae. The objective of this thesis is to examine ways to develop similar algorithms for testing other kind of samplers. In the rst result presented in this thesis we propose a framework to harness Barbarik to test uniformity of state-of-art samplers which are used in the context of Horn formulae. Horn formulae is a subclass of general Boolean formulae and have been quite popular in a variety of domains, due to their simple yet powerful expressiveness, while being able to entail polynomial procedures for solving. In this thesis, we propose an e cient procedure to test samplers for Horn clause, and we demonstrate the e cacy of our testing framework with experiments on large scale benchmarks. In the second part of this thesis, using a di erent form of the conditional sampling model, we investigate the problem of testing samplers in the context of uniform Perfect Matching in a Graph. The problem of nding uniform perfect matching in a graph is of importance in computer science and has a large volume of research associated with it. Modern complex networks often need to nd a uniform perfect matching. The problem of nding a uniform perfect matching nds it's root in physical sciences like statistical mechanics and chemistry. This has inspired research on methods for testing whether the algorithm in use generates a perfect matching uniformly at random. Our contribution in this thesis is a randomized algorithm using the framework of sub-cube conditional sampling to test samplers for uniform perfect matching.
  • No Thumbnail Available
    Item
    Event Timeline Generation and Summarization using PageRank
    (Indian Statistical Institute,Kolkata, 2019-07) Das, Pabani
    Given a search query our system retrieves most significant events related to the query and display them in user friendly timeline. We are using graph structure to compute relative importance of events. We are representing event and its elements as complete graph. To rank events we are linearly combining basic IR ranking and ranking of event elements. evaluation on DUC Data experimental results shows that our system performs at par with the NLP based summarization techniques and yet works as fast as a search engine
  • No Thumbnail Available
    Item
    Feature Extraction And Detection of Malicious URLs Using Deep Learning Approach
    (Indian Statistical Institute,Kolkata, 2019-07) Kushwaha, Rajni
    Phishing Attack is one of the cyber bullying activity over the internet. Most of the phishing websites try to look similar to legitimate websites, their web content and URL features memic the legitimate URL. Due to emerging new techniques, detecting and analyzing these malicious URL is very costly due to their complexities. Traditionally, black and white listing is used for detection, but these technique was not good for real time.To address this, recent years have witnessed several e orts to perform Malicious URL Detection using Machine Learning. The most popular and scalable approaches use lexical properties of the URL string by extracting Bag-of-words like features, followed by applying machine learning models such as SVMs, Randon Forest etc. Various machine learning and deep learning techniques are used to improve generalization of malicious URLs.These approaches su er from several limitations: (i) Inability to e ectively capture semantic meaning and sequential patterns in URL strings; (ii) Requiring substantial manual feature engineering; and (iii) Inability to handle unseen features and generalize to test data. To address these Limitation, In this dissertation work, we are focused to built the real time and language independent phishing detection model by analyzing the anatomy of the URLs using deep learning techniques. To achieve this, we rstly try to nd static and dynamic features manually using some previous work. After getting the featured valued data set, we tried to nd the lexical features of Url using CNN which has both characters and words of the URL String to learn the URL embedding. After that we merge features which we manually selected and features learned from CNN and applied on Bi-LSTM Model to keeps the sequence information of URL. A hybrid model of CNN (convolution neural network model) and Bi-directional LSTM(Long Short Term Memory) are to achieve the goal. Our model analyze the URL without accessing the web content of websites. It eliminates the time latency.
  • No Thumbnail Available
    Item
    A Novel Approach to Automated Coal Petrography Using Deep Neural Networks Souptik
    (Indian Statistical Institute,Kolkata, 2019-07) Mukhopadhyay, Souptik
    This research work is industry sponsored and carried out in collaboration with Tata Steel, India. Its objective is to alleviate a bottleneck in the steel manufacturing pipeline by the application of automated coal petrography. The problem can be de ned as generating semantic segmentation of microscopic coal petrography images. We are presented with a heavily imbalanced and weakly labelled dataset having major intensity based interclass confusion. We have attempted to solve this challenging problem by adopting a deep learning ap- proach to do away with the painful feature engineering process that is often a necessity in classical machine learning. The segmentation task is approached as a pixel level multiclass classi cation problem. Our novel solution uses ve binary U-Net classi ers in accordance with the One-vs-All approach to multiclass classi cation. These binary classi ers are trained using loss functions having additional regularization terms that we have developed in order to handle the interclass confusion problem. These regu- larizers have succesfully resolved majority of this confusion. The result obtained by amalgating the output of the binary classi ers is termed as coarse-segmentation and it su ers from both unclassi ed and misclassi ed pixels. These errors are corrected us- ing a post processing module having four self-developed image processing algorithms and a ne-segmentation is obtained as the nal result. Our solution's performance is benchmarked against two previous approaches based on a Miminum Distance Clas- si er and a Random Forest Classi er. Our method creates superior segmentations that have greater visual appeal and are more accurate. All experimental results are included to support our claim. It was also observed that our results were nearest to those obtained from the current, non-automated standard procedure used in the industry at present.
  • No Thumbnail Available
    Item
    Computational Learning Theory aspects of Piecewise Polynomial and Sigmoidal Neural Networks
    (Indian Statistical Institute,Kolkata, 2019-07) Das, Soumya Kanti
    VC (Vapnik Chervonenkis) Dimension is a useful tool for measuring the power of a neural network or some other types of classi ers. In the eld of learning theory VC dimension represents the generalized power of a neural network. From mid 20th century researchers have been interested in this work and have provided a vast horizon of upper and lower bounds for VC dimension of a neural network. Most of the published work assumes feed forward neural network with no skip connections to establish the upper and lower bounds of VC dimension. In this work we establish that the upper bound of VC Dimension for neural network with piece wise polynomial activation functions can be tighter. Along with this we proposed some other methods for calculating VC Dimension upper bound for RVFLN (neural network with skip connections). Most of the relevant work on VC Dimension upper bound for neural network with sigmoidal activation functions are based on model theoretic approach or number of operations on a basic computing model. Later in this work we give a di erent approach for calculation of VC Dimension upper bound for neural network with sigmoidal activation functions. Moreover on top this we give an idea about how a theoretical test error rate and practical test error rate depend upon on the number of layers and the number of parameters for a feed forward neural network.