Leveraging LLM Generated QA Pairs for Biomedical Question Answering
Developed a novel approach to enhance Biomedical ODQA by generating and using QA pairs from PubMed abstracts.
Improved accuracy by 5% in Factoid and List-type questions compared to document retrieval methods.
Achieved comparable performance in Yes/No and Summary-type questions while reducing inference latency by 15% across all types.
Multi-evidence Natural Language Inference (NLI) for Clinical Trial Data
The project aims at determining the entailment or contradiction given a hypothesis and premise of clinical trials on Breast Cancer which requires numerical and quantitative reasoning of natural language.
Identified effective training strategies for language models in NLI, focusing on textual alignment and evidence retrieval.
Demonstrated that multi-stage fine-tuning significantly improves alignment performance, with task-specific and domain-specific approaches outperforming direct fine-tuning.
Found that MLM-based domain-specific fine-tuning is more effective than DKI for NLI tasks.
Showcased the benefits of context incorporation techniques (e.g., Bag of Words, BM25, Summary) in enhancing evidence retrieval performance.
Developed an end-to-end deep learning model, DOANet, for SSL in drones, leveraging a 1D dilated CNN to compute azimuth and elevation angles directly from raw audio signals.
Improved traditional SSL methods by integrating speed-correlated harmonics cancellation (SCHC) to reduce drone ego-noise.
Demonstrated that DOANet outperforms baseline angular spectrum-based methods, eliminating the need for hand-crafted audio features or additional noise reduction.
Introduced the use of AUC of cumulative histogram plots of angular deviations as a novel performance indicator for SSL evaluation.
Created a dataset from 30 individuals, capturing PPG signals and corresponding glucose levels measured invasively before and after breakfast.
Trained a CNN model to accurately estimate blood glucose levels by exploiting the relationship between PPG signals and glucose concentration.
Achieved promising results, with the system's measurements closely aligning with traditional invasive methods.
Developed a low-cost autonomous robot for collecting trash, addressing the issue of non-biodegradable waste accumulation.
Implemented deep learning algorithms for trash detection, using a camera module and ultrasonic sonar sensors for real-time object classification.
Demonstrated high accuracy in detecting a wide range of trash, contributing to both environmental sustainability and economic efficiency
Estimation of Macronutrients with Continuous Glucose Monitors
The goal of the research was to estimate dietary intake automatically by analyzing the post-prandial glucose response (PPGR) of a meal, as measured with continuous glucose monitors (CGM). Specifically, we wanted to estimate the amount of macronutrients (protein, fat, carbohydrate) from CGM responses using machine learning algorithms.
Gaze for Responsive Interaction with 3D Avatar in Mixed Reality Environment
The goal of the research was to develop a real time system where 3D avatar can respond in a virtual environment based on the gaze to gaze interaction with the user.