Karan Praharaj Email & Phone Number

Chicago, Illinois, United States

- Created a framework for high-privacy continual learning over multilingual data and researched methods to do this optimally. This work culminated in a research paper which was accepted at the Incremental Learning workshop at the International Conference on Data Mining.- Researched the modularization of large NLP models using adapters and achieved a faster training time and a multi-fold improvement in efficiency in terms of model size by only requiring to maintain a small language specific component for each language.- Currently working on a new semantic search functionality which is augmented with document-based question answering.

Chicago, Illinois, United States

Created a framework for continual training of multilingual models with "language-on-top" fine-tuning wherein the model is fine-tuned over different languages incrementally. The developed approach mitigated catastrophic forgetting and improved performance on languages used in fine-tuning, but also on languages not used in fine-tuning.Formulated fine-tuning regimes for multilingual BERT. Achieved an F1 performance improvement of over 23% compared to the previous models in production.Investigated effect of shuffling, random seed, layer-wise learning rate decay, early-stopping, sequential fine-tuning and custom pre-training.

Developed conceptual similarity algorithms for ontology-based similarity evaluation for Wikidata Q-nodes as part of the DARPA KAIROS program.Put it simply, the evaluation was aimed at how conceptually similar predicted entities were to ground truth entities based on a ontology (like Wikidata). For instance, "head of state" is conceptually closer to "head of government" than to "minister". We used our evaluation framework to analyze the accuracy of knowledge-graph based as well as neural embedding methods.

Teaching Assistant for the CSCI 4/5314 Dynamic Models in Biology course offered by Prof. Orit Peleg. The class is cross-listed and was taken by undergrads, MS and PhD students.

Caen Area, France

Developed a novel methodology for semantic similarity and relatedness tasks. The approach incorporates word-neighbors to form "patches" and employs a PageRank attention mechanism to assign scores to neighbors based on centrality. Our inter-patch attention focuses on centrality between patches and enhances the model's performance by incorporating information from the lexical and semantic "neighborhood" of words.

Hyderabad Area, India

Developed an automated academic conference review matching system using statistical ML at the Centre for Visual Information Technology in IIIT Hyderabad. The system leverages Latent Dirichlet Allocation (LDA) to automate the process of assigning submitted research papers to the most suitable reviewer and thereby raise the percentage of expert reviewers by 37%.

Master'S Degree, Computational Linguistics

Bachelor Of Technology, Electronics And Communications Engineering

Icse