Krithika Raj Dorai Raj

• Performing unsupervised clustering (Latent class analysis) of conjoint survey responses to identify latent classes that exhibit similar perceptions of fairness, amongst the population
• Analyze and identify patterns in covariates of the respondents with similar preferences (to study how the covariates might influence perceptions of un/fairness)

• Generated images of forest fires using a curated dataset of web-scrapped images, with a Convolutional Generative Adversarial Network(GAN), as an alternative to the lack of large publicly available, diverse images of forest fires.
• Used the generated images to train a deep convolutional classifier to identify images of forest fires from low resolution, noisy images.

• Causal analysis of conjoint data of the MME (by MIT, choice experiment emulating the trolley problem for decision making in Autonomous Vehicles) to quantify moral preferences
• Hierarchical clustering of country-wise estimates of moral preferences resulted in 3 distinct clusters of preference patterns.

• Implemented prediction models for multi-label classification(one vs rest classifiers) using Apache Spark
• Developed individual models with term document, Tf-Idf and word2vec embeddings as features for prediction and compared their performance.
• Built a prediction model with both word2vec embeddings and Tf-Idf as features to perform the multi-label classification

• Studying how bias manifests in models and the notion of fairness and explainability.
• Implemented a prediction model to identify and evaluate unintended demographic bias following the framework and metric (Relative Negative Sentiment Bias) described in a publication by Chris Sweeney et al.
• Compared the measured bias in word2vec, GloVe and ConceptNet word embeddings against demographic and LGBTQ+ identity terms illustrating that ConceptNet has the least RNSB, followed by GloVe.

• Implemented text processing algorithms from scratch to perform word count, inverted index and n-grams using the Gutenberg dataset on Hadoop framework.

• Built a simulation of an indoor fire-fighter bot (turtlebot3 mounted with OpenManipulator arm) to localize itself in a room using SLAM.
• Navigate to target locations in the room and activate the arm to emulate the act of aiming water jet at the fire till extinguished.