In this project, I am developing both fine-tuned as well as large language models that can incorporate domain expert knowledge, given in the form of key-value pairs, during inference. This methodology helps to tackle the open-vocabulary construct problem where the semantic parser fails to parse the natural language statement into the correct construct in the target domain space if the construct was out-of-distribution with respect to training data. We show the efficacy of our method in both synthetic and real-life settings. In particular, we show how useful our system is when translating operations from the Network File System domain into target formal models such as Linear Temporal Logic. This work is currently under review at EMNLP 2023.

In this project, I develop pool-based active learning methodologies in order to reduce the burden of procuring large amounts of labelled data in order to train supervised models for fake news detection. Combining the common uncertainty based active learning algorithms along with an one-for-all model, I was able to reduce the amount of annotations needed to as low as 4% on certain benchmark datasets. This work has been accepted in the International Conference on Advances in Social Networks Analysis and Mining (ASONAM), 2020 and the link is here 🔗.

In this project, I work on bridging the gap between the biological brain and the statistical neural network. The project involved developing different weight initialization algorithms based upon synaptic distribution inside the mammalian cerebral cortext. Results indicated that our algorithms were good initial weights in comparison to current state-of-the-art and showed positive correlation with model performance. This work has been published in the International Journal of Pattern Recognition and Artificial Intelligence (IJPRAI) and can be found here 🔗.

In this project, I combine the dimensionality reduction algorithm known as Principal Component Analysis along with clustering algorithm based quantization in order to create a methodology that can compress a neural network into a smaller architecture via parameter matrix size reduction. My work was able to compress models upto 35 times on certain benchmark with little to no loss in performance. This work has been published in Applied Intelligence, Springer and can be found here 🔗.

Weakly Supervised Active Learning Scenario for Named Entity Recognition in Difficult Annotation Settings

In this project, I developed a training methodology that combines weak supervision and active learning in a single setting to get more performance with lower amounts of annotated data for the named entity recognition problem. This is done by first training the model on weakly generated labels (labels expressing existence of entity in a statement) and then transfer learning the learnt model to train on fine-grained examples that are optimally selected using active learning methods. The link to the report is here.

Understanding the Variables Affecting Covid-19 Vaccination Rates Across Counties in the United States

This project involved analysis the various elements that are affecting the distribution and immunization of the population against Covid-19 across the different counties in the United States of America. Along with analysing vaccination rates with respect to socio-economic variables, I also observed the sentiment of the tweets from Twitter that are related to the Covid vaccine. In the end, I developed machine learning models that were able to predict future vaccination rates for each of the county. The link to the report is here.

Connect Four Game System Based on LED and Microcontroller

This was a project during the third year of undergraduate education for the "Micro-controller" course. The popular Connect-Four game system is designed using AT-Mega 32 on a screen of 4 LED Matrix display. Controls are implemented using Gyro sensor. The micro-controller based game system is programmed for two players or one player against an AI, developed using the Minimax algorithm. A sample video of the application is available at this link.

Edu Engineer

This was a mobile application developed using Java and Android Studio during my second year as part of an international contest IEEEMadC 2016. It was designed to provide a mobile based platform for sharing information about research and programming,aimed at junior year electrical and computer science students. A sample video of the application is available at this link and the code is here.

Posture Corrector

This was a project during the fourth year of undergraduate education, as part of the "Digital Systems Design" course. Due to rise of backpain as a result of sitting in irregular positions for a long time, we devised an Arduino based system that alarms a person when he/she is not sitting properly. A sample video of the application is available at this link.

Crime Watch

This project is a mobile application developed using the Ionic and Django framework as part of the "Software Design" course during the third year of undergraduate studies. The application enables the user to report and learn about the crimes and various criminal statistics about areas and neighbourhoods close to their location. The code for this project is available at this link.