Projects
A decision-making algorithm based on Federated Co-Learning : Bachelor’s Thesis [Ongoing]- Developing an algorithm to incorporate Co-Training into Federated Learning such that data is decentralized yet benefiting both models mutually.
- Utilizing Liquid Neural Networks as the base model in Federated Learning Network.
Optimizing Unlearning LLMs with advanced neuro-inspired algorithms- Developed and improved a state-of-the-art unlearning algorithm for Large Language Models by adding extensive representations of the memory system, modulations of the attention mechanism, and other cognitive processes motivated by neuroscience.
- Created an adaptive threshold mechanism for the reconsolidation window to optimise unlearning timing and requirement depending on error severity, concept usage frequency, and emotional connotations.
LLM-powered Q&A system using Langchain + Google Palm- Developed an LLM-powered Q&A system (Langchain + Google Palm) within Streamlit UI to reduce workload for e-learning company.
- Spearheaded the implementation of FAISS vector database for fast retrieval of relevant answers, optimizing system performance and user satisfaction.
- This innovative approach enables real-time response and seamless access to knowledge within the e-learning platform.
Framework for adaptive learning in constrained data- Used Bayesian Meta-Learning to design a framework that adapts fast and performs efficiently with limited data.
- Designed a Few-Shot Learning paradigm which achieved 80% accuracy in the classification tasks.
ML framework for Text Classification using Meta-Learning with GANs- Applied a cutting-edge Meta-Learning framework to substantially improve model performance and facilitate rapid adaptation to dynamically changing environments and tasks.
- Developed GAN for data augmentation consisting of a multi-layered generator and discriminator.address Catastrophic Forgetting during Meta-Learning.
- Achieved an accuracy of 95%, compared to 87% for baseline models.
Technique to detect biases in Distributed ML Algorithms- Designed 5 clients and 5 agents system for distributed learning.
- Explored Differential Privacy techniques to add noise to the aggregated model updates, ensuring privacy in Federated Learning.
- Analyzed security in Swarm Learning using Decentralized Identity Management.
Framework for harnessing ML to create Wearable Medical Technology- Carried out clinical studies with the goal of categorising mandibular motions during eating.
- Real-time sensor data acquisition and transmission enabled by the HC-05 Bluetooth Module.
- Compared computational efficiencies of LNN and LSTM for sequential data.
A Transfer Learning framework for detecting vehicles in aerial images- Utilised the potential of transfer learning with VGG16 network architecture.
- Attained quicker convergence on vehicle detection dataset.
- Conduced trials on grey and coloured images and obtained model accuracy of 60%.