Aditya Shah

Mumbai, Ind · EMAIL

I completed my Research-based Master's degree in Computer Science from Virginia Tech where I did MS thesis under Dr. Edward Fox on designing novel NLP methods for preventing diabetes using EFT cues. I completed my Bachelor's degree in Computer Engineering from DJSCE where I served as a Research Assistant under Prof. Lakshmi Kurup on NLP Text Generation. I did my undergraduate thesis on Deep Neural Style Transfer under Prof. Khushali Deulkar.

Previously, I served as a Research Fellow at IIT Indore where I worked with Prof. Chandresh Maurya on designing a novel NLP Multimodal Neural Network architecture for sarcasm detection. I also served as a Deep Learning Researcher at Saarthi.ai where I worked on ASR and Gender Identification from raw audio files.

I've had the pleasure of interning at:

Research Interests:
  • Natural Language Processing, Text Generation, Information Extraction
  • Multimodal Machine Learning, Joint Language-Vision Models, Multilingual NLP
  • Deep Learning, Automatic Speech Recognition, Spoken Language Processing

News

  • Awarded with AI research Fellowship from FELLOWSHIP.AI in May 2020.
  • Research paper on 'Evolution of Neural Text Generation: Comparative Analysis' got accepted for publication in Springer journal at IC4S in Jan 2020.
  • Research paper on 'Leveraging Quantum Computing for Supervised Classification' got accepted in IEEE Xplore at ICICCS 2020 in Nov 2019.
  • Research work on 'Texture Synthesis and Style Transfer for Aesthetic Design Creation' got published with a spotlight in Springer journal at ICACTA in Oct 2019.
  • Delivered a Talk on Quantum Machine Learning, April 2019 at Dwarkadas J. Sanghvi COE followed by a hands-on coding session on Quantum SVM using Qiskit.
  • One of the 200 researchers around the world to get invited to the Qiskit Camp at the IBM T. J. Watson Research Centre, New York, Feb 2019.
  • Founder, Art of Quantum blog, Jan 2019. Published tutorials to explain the key concepts of Quantum Machine Learning using Python and IBM's Qiskit framework.
  • Selected as the youngest Research Scholar at IIT-Bombay, Summer 2018 . Worked under the guidance of Prof Ganesh Ramakrishnan in the field of Machine Learning and Android devlopment.
  • Selected as the Co-Technical Head for Association for Computing Machinery (ACM) 2017-18. Mentored the team for Software Development and Innovation in the fields of Android and web development.
  • Received Google India Scholarship Recipient, 2017. Got hands-on experience on Android Application and Software Development under the guidance and expertise of Google Developer Mentors.

Education

Virginia Tech

Master of Science - Computer Science (Research)

Relevant Coursework: Advanced Machine Learning | Natural Language Processing | Deep Learning | Multimodal Machine Learning

Thesis: Leveraging Transformer Models and Elasticsearch to Help Prevent and Manage Diabetes through EFT Cues Advisor: Dr. Edward Fox
Aug 2021 - May 2023

GPA: 3.9

Dwarkadas J. Sanghvi College of Engineering

Bachelor of Engineering in Computer Engineering
Co-Technical Head - Association for Computing Machinery (2017-2018)

Relevant Coursework: Algorithms | Database Management System | Web Design | Machine Learning | Artificial Intelligence | Natural Language Processing

Thesis: Texture Synthesis and Style Transfer for Apparel Industry     Advisor: Prof. Khushali Deulkar
Teaching Assistant: Machine Learning - Feb 2020
Aug 2016 - Oct 2020

GPA: 9.4 ( In Top 5% )

Experience

Research Fellow - NLP

Indian Institute of Technology (IIT), Indore

Advisor: Prof. Chandresh Maurya

  • Conducting research on developing a novel multimodal neural network architecture for sarcasm detection using text, image, and image attribute input features.
  • Fusing the modal with self matching layers to achieve SOTA results for sarcasm detection.

Sep 2020 - Present

Deep Learning Researcher - NLP and Speech

Saarthi.ai

Saarthi is recognized as one of the Top 20 Conversational AI vendors globally, and also incubated by NVIDIA AI

  • Conducted applied research on ASR and developed a deep learning modal based on LSTM and 1D CNN, achieving SOTA results with a test accuracy of 96% for gender Identification from audio data.
  • Further, worked on age identification and specific keyword detection from the real-time audio input.

Jul 2020 - Oct 2020
(Part time)

Machine Learning Engineer - Computer Vision

QuickFits
  • Led the development of a scalable model for image recognition and real-time visual apparel recommendation on 25 different fashion outfits.
  • Implemented transfer learning on top of Xception model to identify apparels with a test accuracy of 96%.

Mar 2020 - Aug 2020
(Remote Contract)

Research Assistant - NLP

Dwarkadas J. Sanghvi College of Engineering

Advisor: Prof. Lakshmi Kurup

  • Studied Transformers, ELMo, BERT, GPT2, and fine tuned BERT and GPT2 for generating English text samples and evaluated them using BLEU and Self BLEU scores.
  • Developed LSTM model with N-gram and iNLTK library for context-dependent Hindi text generation.

Jan 2019 - Aug 2019

Deep Learning Research Intern - Computer Vision

Fynd Research Lab

Fynd is India’s largest offline to online (O2O) co works with retailers to enable their store inventory to be shoppable across different online and offline channels beyond the store

  • Developed a scalable convolutional neural network (CNN) model using Keras and TensorFlow for dynamic image classification and improved the previous model accuracy by 7%.
  • Contributed in implementing the research paper on “Visual Similarity using Deep CNN” and optimized the model for real-time image similarity.

Sep 2019 - Oct 2019

Machine Learning Intern

Bohr Technology Inc

BOHR∞ (Bohr Technology Inc.) used to develop quantum machine learning software and algorithms for solving complex optimization problems.

  • Studied various research papers on Quantum Support Vector Machine (QSVM), Quantum Neural Networks and Quantum CNN.
  • Implemented ‘Quantum SVM’ using IBM's Qiskit framework to solve binary and multiclass classification tasks at a better accuracy than the classical machine learning approach.

Feb 2019 - May 2019
(Remote)

Research Scholar - ML and SDE

Indian Institute of Technology (IIT), Bombay

Advisor: Prof. Ganesh Ramakrishnan

  • Developed an unsupervised clustering algorithm for a large database of products with XGBoost for "LokaCart", a national award-winning app for farmers; which improved the accuracy by 16%.
  • Designed database schema, front end and back end logic for the android app, and contributed in building MVVM architecture with an asynchronous pool of threads for parallel computing.

Jun 2018 - Aug 2018

Publications

Evolution of Neural Text Generation: Comparative Analysis


Lakshmi Kurup*, Meera Narvekar*, Aditya Shah*, Rahil Sarvaiya* [Paper] [Presentation]

In the past few years, various advancements have been made in Language Models owing to the formulation of various new algorithms such as Generative Adversarial Networks (GANs), ELMo and Bidirectional Encoder Representations from Transformers (BERT). Text Generation, one of the most important Language Modelling problems has shown great promise recently due to the advancement of more efficient and competent context dependent algorithms such as ElMo and BERT and GPT-2 compared to preceding context independent algorithms such as word2vec and GloVe. In this paper we compare the various attempts to Text Generation showcasing the benefits of each in their own unique form.

2020



Texture Synthesis and Style Transfer for Aesthetic Design Creation


Aditya Shah*, Dhruvin Shah*, Harsh Shah, Sneha Shahane, Khushali Deulkar [Paper] [Presentation]

Content rendering of the image by combining different styles has always been a challenging image processing task. The paucity of image representations and content separation from the respective semantic style of the images has been quite an arduous process. However, the recent advancements in image processing, computer vision and state-of-the-art models like VGGNet and ImageNet have helped to achieve human-like accuracy and efficiency in various tasks like generating artistic images, creating design patterns from content images, identifying objects within images, image recognition and segmentation, etc. One such application in this field is deep neural style transfer. It is an optimization technique in which the semantic content of one image is superimposed with the style of another image. Thus, this paper focuses on implementing deep neural style transfer to generate styled images as an application for interior graphic designing and apparel industries. The images generated can be personalized according to the users’ custom design and fashion styles.

2020



Leveraging Quantum Computing for Supervised Classification


Aditya Shah*, Maulik Shah, Pratik Kanani [Paper] [Presentation]

Enhancing quantum computing for supervised machine learning is an innovative application in the field of smart computing. With recent advancements in quantum computing, and it is rising to coalesce with Artificial Intelligence, quantum computers can revolutionize the way to address previously untenable problems. As quantum computers can succeed in producing various intuitive patterns that are strenuous for a classical system to implement, it is reasonable to presume that these quantum machines can outperform a classical computer in various tasks. They can excel at solving problems which involve data crunching with a huge amount of inputs such as machine learning tasks, complex optimization problems, communication system analysis, etc. which require complex parallel computations for an efficient result. This paper attempts to analyze one such aspect of machine learning known as supervised classification with the help of a real Quantum hardware.

2020



* denotes First Author

Projects

Texture Synthesis and Style Transfer for Aesthetic Design Creation

  • Developed a code base for apparel industries where user can get a visual representation for personalized clothes which are styled according to the user preference.
  • Further, implemented Deep Style Transfer by superimposing semantic content of one image with the abstract styled image in order to generate personalized aesthetic images.

Supervised Text Generation using GPT2 model; BiLSTM and GloVe embedding

  • Conducted applied research on GPT2 for Supervised Text Generation and fined tuned GPT2 model on wikisent data for generating context dependant text samples
  • Developed a BiLSTM using GloVe embedding and N-gram model for training on the wikisent dataset to generate context dependant text samples with 89.6% test accuracy

`

Food-101 Challenge by ETH Zurich

  • Conducted applied research on best SOTA models for feature extraction from noisy images.
  • Designed an efficient Neural Network model on top of Xception network and fine-tuned it to achieve State-of-the-Art result on the challenging Food 101 Dataset with a test accuracy of 87%

NLP Recommendation System

  • A robust system which uses content based filtering through word_to_vec encodings and tf_idf / CostVectorizer to recommend various movies using features like movie genre, rating, cast etc.
  • Implemented cosine_similarity score to evaluate the simalarity of words and accurately recommend the 10 most similar movies based on the input title of the movie.

`

Articlex

  • Developed a Full-stack responsive web application using HTML5, CSS3, Bootstrap4, Python, Flask, SQLAlchemy and deployed the app on Heroku.
  • Through this web app, users can create articles and post them, update them, view other posts and can manage their user accounts dynamically.

`

More projects can be found here

Skills

Maths and AI Theory
  • Multivariable Calculus
  • Multivariate Statistics
  • Convex Optimization
  • Probability Theory
  • Linear Algebra

Deep Learning and ML Frameworks
  • TensorFlow
  • Pytorch
  • Keras
  • Pandas
  • NumPy
  • Scikit Learn
  • SpaCy
  • NLTK

Languages and Operating System
  • Python
  • Java
  • JavaScript
  • Windows
  • Linux
  • Ubuntu
  • git

Database Technologies
  • PostgreSQL
  • MongoDB
  • SQLLite
  • MySQL

Software Development
  • HTML5
  • CSS3
  • Bootstrap
  • Flask
  • Heroku

Certifications & MOOC

Stanford University

CS229: Machine Learning
CS231n: Convolutional Neural Networks for Visual Recognition
CS224n: Natural Language Processing with Deep Learning

DeepLearning.ai

Tensorflow in Practice

Deep Learning Specialization