K. K. Shukla | IIT BHU (original) (raw)

Papers by K. K. Shukla

arXiv (Cornell University), Dec 5, 2020

The ability to interpret decisions taken by Machine Learning (ML) models is fundamental to encour... more The ability to interpret decisions taken by Machine Learning (ML) models is fundamental to encourage trust and reliability in different practical applications. Recent interpretation strategies focus on human understanding of the underlying decision mechanisms of the complex ML models. However, these strategies are restricted by the subjective biases of humans. To dissociate from such human biases, we propose an interpretation-by-distillation formulation that is defined relative to other ML models. We generalize the distillation technique for quantifying interpretability, using an information-theoretic perspective, removing the role of ground-truth from the definition of interpretability. Our work defines the entropy of supervised classification models, providing bounds on the entropy of Piece-Wise Linear Neural Networks (PWLNs), along with the first theoretical bounds on the interpretability of PWLNs. We evaluate our proposed framework on the MNIST, Fashion-MNIST and Stanford40 datasets and demonstrate the applicability of the proposed theoretical framework in different supervised classification scenarios.

2021 International Conference on Computing, Communication, and Intelligent Systems (ICCCIS), 2021

Machine Learning models are being increasingly deployed to tackle real-world problems in various ... more Machine Learning models are being increasingly deployed to tackle real-world problems in various domains like healthcare, crime and education among many others. However, most of the models are practically "black-boxes": although they may provide accurate results, they are unable to provide any conclusive reasoning for those results. In order for these decisions to be trusted, they must be explainable. Explainable AI, or XAI refers to methods and techniques in the application of AI such that the results of the solution are understandable by human experts. This paper focuses on the task of Image Captioning, and tries to employ XAI techniques such as LIME (Local Interpretable Model-Agnostic Explanations) to explain the predictions of complex image captioning models. It visually depicts the part of the image corresponding to a particular word in the caption, thereby justifying why the model predicted that word.

International Journal of Computer Applications, 2015

Computer Application and Signal Processing 2014, 2014

Biomedical and Pharmacology Journal, 2017

This paper presents the automated detection and classification of histopathological images of can... more This paper presents the automated detection and classification of histopathological images of cancer cells using morphological features. The manual assessment of disease is time-consuming and varies with the perception and the level of expertise of the pathologists. The judgment is based on the tissue structures, distribution of cells in tissue and the irregularities of cell shape and size. To overcome the limitation of manual diagnosis, a computer aided diagnosis based on the morphological features has been implemented for accurate and reliable detection of cancer. A dataset of 70 histopathological images of non-cancerous and cancerous tissues are randomly selected. The proposed work aims at developing the technique that uses reliable quantitative measures for providing objective and reproducible information complementary to that of a pathologist.

Procedia Computer Science, 2015

Integer Factorization is a vital number theoretic problem frequently finding application in publi... more Integer Factorization is a vital number theoretic problem frequently finding application in public-key cryptography like RSA encryption systems, and other areas like Fourier transform algorithm. The problem is computationally intractable because it is a one-way mathematical function. Due to its computational infeasibility, it is extremely hard to find the prime factors of a semi prime number generated from two randomly chosen similar sized prime numbers. There has been a recently growing interest in the community with regards to evolutionary computation and other alternative approaches to solving this problem as an optimization task. However, the results still seem to be very rudimentary in natur to be done. This paper emphasizes on such approaches and presents a critic study in details. The paper puts forth criticism and ideas in this aspect.

Public key cryptography is extensively used for encryption, signing contracts and secure exchange... more Public key cryptography is extensively used for encryption, signing contracts and secure exchanges over the unreliable network. The findings of Shor in 1994, of a powerful algorithm which was based on quantum mechanics for computing discrete logarithms and factoring large integers sabotaged the security presumptions upon which the currently used public key cryptographic protocols are based, like ElGamal, RSA and ECC. However, few cryptosystems, known as post quantum cryptosystems, while not currently in wide use are considered to be resistant to such attacks. In this paper, a quantum version of ElGamal Cryptosystem is proposed whose security relies on the commutative rotation transformations and measurements in computational basis of qubits. An understanding of the new scheme over the quantum channels is provided. The proposed cryptosystem allows the party to send messages in the form of qubits over a quantum channel. Also the proposed protocol provides an opportunity for two partie...

Nature-inspired computations are commonly recognized optimization techniques that provide optimal... more Nature-inspired computations are commonly recognized optimization techniques that provide optimal solutions to a wide spectrum of computational problems. This paper presents a brief overview of current topics in the field of nature-inspired computation along with their most recent applications in deep learning to identify open challenges concerning the most relevant areas. In addition, we highlight some recent hybridization methods of nature-inspired computation used to optimize the hyper-parameters and architectures of a deep learning framework. Future research as well as prospective deep learning issues are also presented.

Applied Intelligence, 2017

Proximal Algorithms are known to be very popular in the area of signal processing, image reconstr... more Proximal Algorithms are known to be very popular in the area of signal processing, image reconstruction, variational inequality and convex optimization due to their small iteration costs and applicability to the non-smooth optimization problems. Various real-world machine learning problems have been solved utilizing the non-smooth convex loss minimization framework, and a recent trend is to design new accelerated algorithms to solve such frameworks efficiently. In this paper, we propose a novel viscosity-based accelerated gradient algorithm (VAGA), that utilizes the concept of viscosity approximation method of fixed point theory for solving the learning problems. We discuss the boundedness of the sequence generated by this iterative algorithm and prove the strong convergence of the algorithm under the few specific conditions. To test the practical performance of the algorithm on real-world problems, we applied it to solve the regularized multitask regression problem with sparsity-inducing regularizers. We present the detailed comparative analysis of our algorithm with few traditional proximal algorithms on three real benchmark multitask regression datasets. We also apply the proposed algorithm to the task of joint splice-site recognition problem of bio-informatics. The improved results demonstrate the efficacy of our algorithm over state-of-the-art proximal gradient descent algorithms. To the best of our knowledge, it is the first time that a viscosity-based iterative algorithm is applied to solve the real world problem of regression and recognition.

Knowledge and Information Systems, 2017

We consider the problem of minimization of the sum of two convex functions, one of which is a smo... more We consider the problem of minimization of the sum of two convex functions, one of which is a smooth function, while another one may be a nonsmooth function. Many high-dimensional learning problems (classification/regression) can be designed using such frameworks, which can be efficiently solved with the help of first-order proximal-based methods. Due to slow convergence of traditional proximal methods, a recent trend is to introduce acceleration to such methods, which increases the speed of convergence. Such proximal gradient methods belong to a wider class of the forward–backward algorithms, which mathematically can be interpreted as fixed-point iterative schemes. In this paper, we design few new proximal gradient methods corresponding to few state-of-the-art fixed-point iterative schemes and compare their performances on the regression problem. In addition, we propose a new accelerated proximal gradient algorithm, which outperforms earlier traditional methods in terms of convergence speed and regression error. To demonstrate the applicability of our method, we conducted experiments for the problem of regression with several publicly available high-dimensional real datasets taken from different application domains. Empirical results exhibit that the proposed method outperforms the previous methods in terms of convergence, accuracy, and objective function values.

2016 4th International Symposium on Computational and Business Intelligence (ISCBI), 2016

Ensembling classifiers has been an effective technique for improving performance and generalizabi... more Ensembling classifiers has been an effective technique for improving performance and generalizability of classification tasks. In a recent research direction, the ensemble of the random projections is being utilized as an effective regularization technique with linear discriminant classifiers. However the framework has only been designed for binary classifiers. In this paper we extend the idea for the multiclass classifiers, which directly improves the applicability of the framework to a broader class of problems. We performed experiments with multiple high-dimensional benchmark datasets, and compare the performance of our framework with other state-of-the-art methods for multi-class classification. We also extend the theoretical error bounds for misclassification to provide a theoretical analysis. Results demonstrate the efficacy of our methodology.

2016 IEEE International Conference on Systems, Man, and Cybernetics (SMC), 2016

Multitask learning methods facilitate learning multiple related tasks together and improvise resu... more Multitask learning methods facilitate learning multiple related tasks together and improvise results as compared to the schemes where each task is considered independently. In order to incorporate the shared information in multiple tasks, various regularizers have been integrated in pre-existing techniques. In this paper, we explore the problem of convex formulations of multitask learning with sparsity inducing regularizers. The main contribution of this paper is to introduce a novel first order iterative procedure (MTL-FIBM) which we prove to converge faster than previously existing work. Our method belongs to the class of proximal gradient-based techniques, where the loss function is considered to be smooth and the regularization function maybe non-smooth. We performed extensive experiments with synthetic as well as two real datasets namely School and Parkinson Telemonitoring datasets and show that the experimental results agree with the theoretical analysis of our algorithm. Results demonstrate the efficacy and improvement in terms of speed and accuracy.

IEEE Transactions on Fuzzy Systems, 2016

In this paper, a new definition for Atanassov intuitionistic fuzzy metric space is presented usin... more In this paper, a new definition for Atanassov intuitionistic fuzzy metric space is presented using the concept of Atanassov intuitionistic fuzzy point and Atanassov intuitionistic fuzzy scalars. The distance metric introduced here is then applied to an interesting problem called the orienteering problem that finds application in several industries, such as the home delivery system, robot path planning, tourism industry etc., and in each of these practical applications, the two parameters involved, i.e., score and distance travelled as well as the position of locations cannot be predicted precisely. To tackle these uncertainties, we use trapezoidal Atanassov intuitionistic fuzzy numbers for representing the parameter score. The uniqueness of this paper is the consideration of uncertainty in the position of a city or a location and handling this type of uncertainty using the idea of Atanassov intuitionistic fuzzy points and the distance metric between Atanassov intuitionistic fuzzy points. Further, a method for ranking trapezoidal Atanassov intuitionistic fuzzy numbers has been presented and used for modeling the scores.

2014 International Conference on Soft Computing and Machine Intelligence, 2014

ABSTRACT In this paper, the intractable problem of finding the prime factors of an integer has be... more ABSTRACT In this paper, the intractable problem of finding the prime factors of an integer has been epresented as the integer programming problem to be solved using nature inspired heuristics. Integer factorization is a one-way mathematical function and because of its computational intractability, it is frequently used in public key cryptography, for example in RSAencryption systems. Since integer factorization can be represented in the form of a discrete optimization task, more specifically as integer programming problem, various optimization tools can be utilized for cryptanalysis. In this contribution, we approach this problem using a heuristic algorithm designed with concepts from computational chemistry that involves energy minimization of a molecular geometry of a crystal. We observe that computational chemistry can provide a great insight into such problems of unknown dynamics. Future work remains to optimize the algorithm for scalability.

Clustering has an extensive and long history in a variety of scientific fields. Several recent st... more Clustering has an extensive and long history in a variety of scientific fields. Several recent studies of complex networks have suggested that the clustering analysis on networks has been an emerging research issue in data mining due to its variety of applications. Many graph clustering algorithms have been proposed in recent past, however, this clustering approach remains a challenging problem to solve real-world situation. In this work, we propose an aspiration criteria based graph clustering algorithm using stochastic local search for generating lower cost clustering results in terms of robustness and optimality for real-world complex network problems. In our proposed algorithm, all moves are meaningful and effective during the whole clustering process which indicates that moves are only accepted if the target node has neighbouring nodes in the destination cluster (moves to an empty cluster are the only exception to this instruction). An adaptive approach in our method is in incorporating the aspiration criteria for the best move (lower-cost changes) selection when the best non-tabu move involvements much higher cost compared to a tabued move then the tabued move is permitted otherwise the best non-tabu move is acceptable. Extensive experimentation with synthetic and real power-law distribution benchmark datasets show that our algorithm outperforms state-of-the-art graph clustering techniques on the basis of cost of clustering, cluster size, normalized mutual information (NMI) and modularity index of clustering results.

SpringerBriefs in Computer Science, 2013

Discrete wavelet transforms (DWTs) have excellent energy compaction characteristics and are able ... more Discrete wavelet transforms (DWTs) have excellent energy compaction characteristics and are able to provide near perfect reconstruction (PR). They are ideal for signal/image analysis and encoding. Hardware implementation of DWT is fast and area efficient in fixed-point arithmetic. DWT encoding has been drawing much attention because of its ability to decompose signals into a hierarchical structure that is suitable for adaptive processing in the transform domain. In existing architectural designs for the DWT, little consideration has been given to word size and precision. Present chapter addresses this problem, showing how the word size requirements can be calculated for a specific problem (based on the range of input data and wavelet used). A simplified, statistical model is proposed. As the depth of the DWT filtering increases, the data word length requirement increases. It is important to investigate how this can affect the potential of the resulting hardware implementation of DWT. The issue has been analyzed for both pyramid structure DWT and parallel filter DWT. The organization of this chapter is as follows. Section 3.1 presents background material related to subject. Section 3.2 presents in brief the computational complexity of DWT. Section 3.3 presents finite precision modeling of two-channel PR filter bank in moderate detail, including modeling of quantized coefficient filters. Section 3.4 presents the proposed statistical modeling of DWT to study the effects of finite word length implementation. This includes construction of new DWT filters to accommodate round-off errors followed by corresponding mathematical derivation.

IoT-Based Data Analytics for the Healthcare Industry

Predicting the number of bugs in any software application is an important but challenging task. T... more Predicting the number of bugs in any software application is an important but challenging task. The software manager by modelling the bug numbers, can take timely decisions in reducing the amount of effort investment and also the allocation of resources. The software developers can also take effective steps for reducing the number of bugs in the future version of the software application. The end users also can make a timely decision on adoption of a particular software application by knowing the growth pattern of bugs in advance. The challenges behind modeling the bug growth patterns are random causes behind a bug. A bug in any software may be caused during testing, development or application. Causal modelling of bug numbers is a complex and tedious task as they consider many internal characteristics to be modelled. In this paper, we have used we have used Long Short Term Memory (LSTM) [14] Network for temporal modelling the bug numbers of three different software applications. We ...

The International Journal of the Computer, the Internet and Management, 2005

Task scheduling is very important in Real-Time Systems as it accomplishes the crucial goal of dev... more Task scheduling is very important in Real-Time Systems as it accomplishes the crucial goal of devising a feasible schedule of the tasks. However, the uncertainty associated with the timing constraints of the real time tasks makes the scheduling problem difficult to formulate. This motivates the use of fuzzy numbers to model task deadlines and completion times. This paper reports how the changes in the membership functions of the fuzzy deadlines affect the satisfaction intervals, the satisfaction of schedulability and the tasks priorities of set of a real time tasks.