Ashrya Agrawal is a Data Science Intern at JP Morgan. He is also a undergraduate Computer Science student at BITS Pilani, Pilani Campus. He is passionate about mitigating bias in Machine Learning models. His journey of fairness began with JSoC 2020 where he developed a toolkit for bias audit and mitigation.
Research in Fairness. While working on the debiasing algorithms in the JSoC toolkit, he realized that the debiasign algorithms often fail in practice to generalize out-of-sample. This piqued his interest, and he was curious to understand these failures. In his research he demonstrated that the failure of debiasers can be explained as a consequence of large variance due to the bias-variance trade-off. To mitigate this large variance and improve generalization, he formulated the notion of partial debiasing.
Research in Causality. After studying debiasing from an observational viewpoint, he studied it from a causal perspective to analyze the behavior of debiasing methods and draw deeper insights than what is provided by statistical analysis of fairness. He studied the effect of debiasing methods on causal path-specific fairness metrics. Then he explained these effects from an Information-theoretic perspective using partial information decomposition.
Apart from fairness and causality, he is interested in exploring promising research areas in Machine Learning. He has worked in ADAPT Lab on Video Captioning using knowledge bases. Ashrya is also passionate about ML reading groups, open collaborations and teaching.
B.E. Computer Science, 2022
BITS Pilani - Pilani Campus
Minor - Data Science, 2022
BITS Pilani - Pilani Campus
Created Friend Affinity Finder which uses social media data to perform NLP analysis.
Recipient of the prestigious KVPY fellowship twice.
We built a web app which shows you the list of various people around you who are willing to combine the orders for e-shopping in order to save delivery charges.
Developed a voice assistant with automatic Object Detection to aid navigation, memory issues for Alzheimer patients using ResNet, cloud, NodeJS, etc. Helps patients by locating objects misplaced or forgotten by patient.
Developed an automatic FAQ-answering system to improve response time for queries by employees. Fine-tuned BERT and computed feature representation using the fine-tuned BERT model
Built a compiler in C, supporting various data-types, arrays, expressions and specifically jagged arrays. Implemented lexical analyser, parser, abstract syntax tree generator and type checker.
Developed Backend for Checkmate 2019, a bi-annual puzzle-based event by BITS ACM.
Video captioning is a challenging task of modelling the objects, their temporal information and interaction in order to generate a textual description. Current models often fail to model these objects and their interactions correctly, due to lack of knowledge about them.
We present an empirical study of debiasing methods for classifiers, showing that debiasers often fail in practice to generalize out-of-sample, and can in fact make fairness worse rather than better.
We study the effects of different debiasing methods on the underlying causal path specific effects (PSEs) and explain these effects using an information-theoretic perspective.
We present Fairness.jl, a comprehensive bias audit and mitigation toolkit in julia. Extensive support and functionality provided by MLJ.jl has been used in this package.
Created a Friend Affinity Finder that performs NLP analysis of the social media profiles of friends, generates affinity score for friends and clusters them.