Research Experience

Google Summer of Code @ Machine Learning for Science (ML4SCI) (May 2023 - September 2023)    

"FASEROH : Building seq2seq model for mapping histograms to empirical symbolic representations" (Florida State University, University of Kentucky)

1. The problem involves creating a seq2seq model for mapping histograms to empirical function sequences. 

2. Step 1 involved creating random functions and normalizing them to get a set of function sequences, then generating histogram data and best-fit parameters for the function. This dataset is proposed to have a million examples. 

3. Step 2 involved experimenting with different seq2seq models like LSTMs, Transformers, and Pre-trained Transformers and training/evaluating their performance for our task.

Senior Thesis, NIT Hamirpur (September 2022 - April 2023)    

"Evaluation of Gender-Bias in Hindi-English NMT Models; Assessing various Social Biases in Multilingual NLP for Indian Context"

1. I constructed two novel test sets in Hindi with grammatical gender markers to evaluate whether NMT models can capture the correct gender information from these gender cues, instead of relying on biased correlations associated with occupations. The study revealed that this context-based evaluation is better at exposing the gender bias of an NMT system otherwise hidden if only measured using gender-neutral sentences. (Extrinsic evaluation)

2. Further worked on the intrinsic evaluation of social biases in various word embeddings (monolingual or multilingual) for the Indian context. 

Associated work: Gender Inflected or Bias Inflicted: On Using Grammatical Gender Cues for Bias Evaluation in Machine Translation (AACL-IJCNLP 2023 SRW, best paper)

University of Southern California, CA, USA (May 2022 - July 2022)    

1. • IUSSTF-Viterbi Summer Research Intern

   • Project Name : "Explanation based Regularisation of Language Models"

   • Advisor: Dr. Xiang Ren (INK Lab)

   • Explanation-based Regularisation deals with using human and machine explanations to regularize the language model to enhance its performance and generalisability across different NLP tasks. Explanation regularisation loss (L1 Loss) is calculated between machine rationales and human rationales. My work involved working with human rationales and coming up with better ways to improve their information. I created and experimented with different versions of human rationales - insertion and deletion rationales. These are obtained from originally available human rationales but essentially, simulate exclusive insertion and deletion operations on human rationales. I also experimented with different loss functions to improve learning from explanations.

Indian Institute of Technology(IIT), Roorkee (May 2021 - July 2021)            

• • SPARK Programme Summer Research Intern

  • Project : "ML Driven State-of-Health Estimation of Li-Ion Batteries" 

  • Advisor: Dr. Ashwini Kumar Sharma

  • For the battery life estimation, I used the discharge capacity dataset from the initial discharge cycles available in form of a time series. Since the training data was limited, it required me to rely more on traditional machine learning approaches including feature extraction, filtering, dimensionality reduction, and using basic algorithms like Linear Regression, Elastic Net, and Gradient Boosting. In one of my experiments, I used a library called ``tsfresh" for automatic feature extraction which yielded some interesting features from a theoretical perspective. Our single-feature model outperformed the ``Variance" model proposed by Severson et al. 2019 on Mean Average Percentage Error(MAPE).

IISER Kolkata, India (June 2021 - July 2021)            

• Summer Research Intern

• Project : "Question-Answering System for Indian Legal Queries" 

• Advisor: Dr. Kripabandhu Ghosh

• I trained traditional vector space models like TF-IDF, word2vec, GloVe, etc, on a web-scraped legal dataset to retrieve the closest answers for a given query. I evaluated the retrieval performance using metrics like Mean Average Precision. In addition to the unsupervised approach, I also used the Pointwise Learning to Rank(Supervised) approach by hand engineering 152 features from the Q-A pairs (after preprocessing) and then applying feature selection(using Recursive Feature Elimination - Cross Validation (RFE-CV)) and finally, training those selected features using algorithms like Support Vector Classifier, Random Forest Classifier, Decision Tree Classifier, Gaussian Naive Bayes and Gradient Boosting Classifier which outperformed unsupervised method.