Debrup Das
Debrup Das, IIT Kharagpur dasdebruprohon20@gmail.com
debrupdas@iitkgp.ac.in

 +918478033850

  Github

I am a 5th year student of Mathematics and Computing at the Indian Institute of Technology Kharagpur, enrolled in an Integrated Master of Science 5 yr degree.
I have been working in Tr^2AL Lab under Prof Somak Aditya, in the domain of neuro-symbolic conversational AI systems.

My research interests are in the domain of NLP with specific interests in:

  • Large language models and their ability to perform mathematical and commonsense reasoning.
  • Neuro-symbolic methods using structured knowledge to alleviate limitations in LLMs.
  • Investigation and mitigation of bias and toxicity in LLMs (Prompt Injection Attacks).

Neuro-symbolic AI
under a probabilistic logical reasoning framework
(Collaborative Project with Deepanaway Ghosal SUTD)

  • Use knowledge graphs to extract information.
  • Represent knowledge using probabilistic logic programs.
  • Working currently with FOLIO and related datasets.

LOGIBOT
(work ongoing with Rakuten Corporation)

  • Compositional methods to improve mathematical reasoning of LLMs on the MATH dataset

Prompt Injection Attacks and Liability
(under the Microsoft Academic Research For Accelerating Foundation Models Grant)

  • Framework for quantifiying the sensitivity of LLMs to different parts of a prompt for jailbreak.
  • Quantifiying the bias of OpenAI models to certain words and use these features for mitigation.
  • Human and automatic evaluation of the framework.

Genealogy GWAS
(under Prof Simon Gravel, McGill University, MITACS 2022)

  • Improving Genome-Wide Association Studies by leveraging a genealogy of 3 million individuals in Lac-Saint-Jean region of Quebec.
  • Using Monte Carlo based simulations of transmission histories of alleles conditioned on existing genealogy and available genetic data.

Deep Forests
(under Dr Nirupam Chakraborti, Czech Technical University, 2021)

  • Perform deep learning using subnets of trees.
  • Using nature inspired Predator-prey algorithm and genetic operators to explore the search space of architectures.
  • Perform multi-objective optimization on both the size of model and RMSE error.

Other Areas Of Interest

I am a keen follower of research in Computer Vision, Genetic Algorithms and other subdomains of AI. I am passionate about statistics and its applications to diverse fields of study. I am always motivated to conduct research in interdisciplinary laboratory environments and explore new fields of study.