Anshumaan Chauhan

Research Interests🔎

I am currently interested and researching on the following topics pertaining to Deep Learning:

Publications 📝

Constraint-Based Multi-organ Identification in CT Images using Unsupervised Learning

Accurate organ identification and segmentation in whole-body Computed Tomography (CT) images are important for the localization of diseases and treatment planning. Manual organ identification and segmentation requires human intervention/expertise and is tedious. Many automated approaches based on supervised machine learning are being used to accomplish scalability, but it requires a large amount of training data that is not easy to access in the case of medical images due to the expense of acquisition and data anonymization. We present a fully automated unsupervised knowledge-based workflow for CT organ segmentation and localization. The workflow consists of pre-processing with ImageJ, followed by segmentation using Density-Based Spatial Clustering of Applications with Noise (DBSCAN), morphological closing operations for post-processing, and a knowledge-based analysis to identify organs. The knowledge-based framework utilizes constraint-based area check (CBAC), conditional localized overlap check, and multi-scale template matching (MTM) using a single template to rule out infeasible segmentations and recognize the organs precisely. The model was tested on an in-house whole-body CT scans dataset comprising 150 patients. The proposed methodology yielded significantly high average Dice Coefficients of 0.784 and 0.883 for kidneys and lungs, respectively, against semi-manually segmented organs by an expert.

Toward Generating System Architecture and Formal Functional Description in the Architecture Analysis & Design Language (AADL) with Structured Natural Language

Model based engineering has enabled automated analytical reasoning early in the design phase. As a result, inconsistencies and design errors can be captured early in the development lifecycle. But there is still a gap in the natural language-based specifications and its actual implementation. This is because the formal method-based tools utilize mathematical principles and theories of computation that require specific skills, thus reducing the usability of model-based engineering. Natural language is the most widely used method to represent specifications. So, it is intuitive to utilize natural language-based representation to generate system and formal annotations such that it will enable automated architectural analysis with much wider acceptance leading to a much broader impact. In our paper we focus on designing the above-mentioned approach that integrates representation of the specifications in a subset of English language which can then be used to generate system architecture in Architecture Analysis and Design Language along with the generation of functional specifications. We illustrate our approach by validating it with use cases from the aerospace and electromechanical domains.

DQNAS: Neural Architecture Search using Reinforcement Learning

Convolutional Neural Networks have been used in a variety of image related applications after their rise in popularity due to ImageNet competition. Convolutional Neural Networks have shown remarkable results in applications including face recognition, moving target detection and tracking, classification of food based on the calorie content and many more. Designing of Convolutional Neural Networks requires experts having a cross domain knowledge and it is laborious, which requires a lot of time for testing different values for different hyperparameter along with the consideration of different configurations of existing architectures. Neural Architecture Search is an automated way of generating Neural Network architectures which saves researchers from all the brute-force testing trouble, but with the drawback of consuming a lot of computational resources for a prolonged period. In this paper, we propose an automated Neural Architecture Search framework DQNAS, guided by the principles of Reinforcement Learning along with One-shot Training which aims to generate neural network architectures that show superior performance and have minimum scalability problem.

LPRNet: A Novel Approach for Novelty Detection in Networking Packets

Novelty Detection is a task of recognition of abnormal data points within a given system. Recently, this task has been performed using Deep Learning Autoencoders, but they face several drawbacks which include the problem of identity mapping, adversarial perturbations and optimization algorithms. In this paper, we have proposed a novel approach LPRNet, a Denoising Autoencoder which uses algorithms such as Least Trimmed Square, Projected Gradient Descent and Robust Principal Component Analysis, to solve the above-mentioned problems. LRPNet is then trained and tested on NSL-KDD dataset, and experiments have been performed using Accuracy as performance metric for comparing the existing models with the proposed model. The results show that LRPNet has the maximum accuracy of 95.9% and performed better than all the previous state-of-the-art algorithms.

Genetic Algorithm and Ensemble Learning Aided Text Classification using Support Vector Machines

Text classification is one of the areas where machine learning algorithms are used. The size of the dataset and the methods used for converting the textual words into vectors play a major role in classifying them. This paper proposes a heuristic based approach to classify the documents using Genetic Algorithm aided Support Vector Machines (SVM) and Ensemble Learning approach. The real valued representation of the textual data into vectors is done on applying Term Frequency – Inverse Document Frequency (TF-IDF) and Bi-Normal Separation (BNS). However, in this paper, the common data misclassification issue in SVM is overcome by introducing two algorithms that adds weightage to accurate classification. The first algorithm applied BNS and TF-IDF along with ensemble learning and constructs a voting classifier for classifying the textual documents. The results produced justify that TF-IDF produces good results with voting classifier than BNS for classification. Henceforth TF-IDF is applied in the subsequent approach for vector generation. Secondly, genetic algorithm is applied along with OneVsRest strategy in SVM to overcome the drawback of multiclass multilabel classification. The results show that Genetic algorithm improves the accuracy of classification even with a very small labelled dataset, as genetic algorithm applies the process of Mutation and Cross over across many generations to understand the pattern of right classification.

Performance Analysis of Machine Learning Algorithms and Feature Extraction Methods for Sentiment Analysis

Sentiment Analysis is the process of evaluating the document or sentence and assigning it a polarity. It has been a key area of research for the past few years. With the evolution of the World Wide Web, many platforms such as Twitter, Facebook, etc. came up where people can express their emotions related to an object, movie, or any political party. These reviews are read by many people before taking some decision, and hence it is very important for the Sentiment Analysis models to assign polarity to the reviews properly. In this paper, we will be analyzing different existing Machine Learning algorithms such as Linear Regression, Support Vector Machine, Decision Trees, Random Forest, and Maximum Entropy Model used for Sentiment Analysis, along with 2 most used methods of feature extraction, Bag-of-Words(BOW) and Term Frequency- Inverse Document Frequency (TF-IDF). The results showed that BOW used with Linear Regression models shows the best results achieving an accuracy score of 34.83% and takes minimum time for training.

Please refer to my Google Scholar for more details about the publications I have been a part of.