Krishnakant Bhatt

About

I am a student at the CSE department of IIT Bombay, currently in the final year of my masters degree. I am enthusiastic about Machine Learning, Large Language Models, Data Structures and Algorithms.

More About Me

I am currently specialized in the domain of NLP and doing my work on Machine Translation in domain specific scenario and Bilingual Lexicon Generation.

Email: kkbhatt@cse.iitb.ac.in
Degree: Masters
Phone: Not Available
Hometown: Neemuch

I am also quite passionate about Physics, Cosmos, Astronomy. Books like "What If?: Serious Scientific Answers to Absurd Hypothetical Questions" by Randall Munroe and YouTube channels like 3B1B, AI Warehouse, Veritasium, The Action Lab and Engineering World aligns perfectly with my interests.

Grades and Scores

Important academic grades and competitive exam scores.

9.03

CGPA (IIT Bombay)

99.1

Percentile (GATE 2021)

Skills

My skillset includes

Python 100%

C/C++ 90%

Git 75%

Bash 60%

PyTorch 90%

Huggingface 90%

NumPy 90%

Pandas 75%

Projects

Following are some of the my important ongoing and completed projects.

Masters Thesis

Lexically Constrained Neural Machine Translation

2022 - Present

Experimented with existing and novel techniques to build a constrained multilingual translation model for Project UDAAN.
Project UDAAN aims to break the language barrier in education and is adopted by the Govt. of Maharashtra.
Developed different approaches to build a deep-learning model for dictionary generation.

R&D and Seminar

Post Editing in MT (R&D Project)

July'21 - Dec'21

Experimented with techniques like constrained decoding and Levenshtein Transformer to rectify translation.
Elevated bank document translation for our industry sponsor Bank of Baroda, achieving BLEU score of 48.1

Seminar

Jan'22 - Apr'22

Performed literature survey and delved into the recent advancements in constrained MT.
Thoroughly examined and classified diverse methods into distinct categories based on the underlying algorithm.

Course Projects

Adaptively Optimizing Subnetworks While Finetuning Pretrained LMs

Jan'23 - Apr'23

Experimented with a NeurIPS-22 approach Dynamic Parameter Selection on the task of sentiment classification.
Examined models like BERT and BART on the Sentiment-140 dataset and analysed their performances.
Achieved an increase of 1.15x and 1.044x in F1-Score using DPS as compared to vanilla DistilBERT and BART.

Improving Adversarial Transferability by Combining Momentum and LinBP

Jan'22 - Apr'22

Experimented with a combination of momentum-based technique and linear backpropagation-based technique to increase the transferability of adversarial examples on Generative Adversarial Networks (GANs).
Achieved improvement of 0.65% in success rate over best baseline on the ResNeXt model using CIFAR-10 dataset.

Question Answering on SQUAD

Jan'22 - Apr'22

Implemented a question-answering system using an RNN-based encoder-decoder.
Achieved an F1-Score of 0.77 on the Stanford Question Answering Dataset (SQUAD).