projects
RAG Optimizations with Reinforcement Learning
Year: Feb - May 2024
Category: Deep Learning, Reinforcement Learning
This project proposes a novel reinforcement learning based optimization strategy that aims to rectify these impediments by evaluating retriever performance on diverse conversa tional data with specific regards to the type of text under consideration. Through the integration of text analysis metrics, different types of retrieval techniques, evaluation frameworks for context and response generation and Reinforcement Learning (RL) policies such as DQN, PPO and Multi-armed Bandit, we propose an evolved RAG framework capable of producing superior responses with a marked reduction in superfluous content.
Khalasi-IO: Detecting Situational Impairements with Reasoning Based Memory Bank
Year: Oct - Dec 2024
Category: Deep Learning, Human Computer Interaction
Detects and addresses temporary impairments caused by environmental factors like noise, lighting, and stress. Integrates wearable devices and contextual memory for personalized, real-time interventions, reducing cognitive load and frustration. Promising results highlight scalability and improved accessibility.
Visual Embeddings Solving Jigsaw Puzzles
Year: Feb - Apr 2024
Category: Deep Learning, Graph Neural Networks
Devised a novel deep learning model architecture and corresponding loss functions dedicated to representation learning for downstream tasks through the pre-text application of jigsaw puzzle paradigms, utilizing Graph Neural Networks and Autoencoders.
Pioneered a Segmented Flow warp approach to enhance representation learning, achieving a validation accuracy of 60%, competing with current state-of-the-art approaches.
Technologies used: PyTorch, CNN, ViGNN, Vision Transformer.
Quantum Machine Learning Classifiers
Year: Mar - May 2024
Category: Quantum Machine Learning
Implemented 24 combinations of Feature Maps, Ansatzs, and Optimizers to research the efficacy of Quantum kernel‐based classifiers.
Our combination of the Z‐Feature map, EfficientSU2 ansatz, and L_BFGS_B optimizer outperformed classical SVM in all metrics.
Technologies used: IBM Qiskit, Quantum Kernel Methods.
SWaV-Based Clustering for Vehicle Identification
Year: Oct - Nov 2023
Category: Self-Supervised Learning, Computer Vision
Used SWaV (Swapping Assignments between Views) based clustering to perform vehicle re-identification on the Veri-776 dataset.
The model assigns the same cluster to the same vehicle across different images and assigns different clusters to different vehicles, effectively improving vehicle identification accuracy.
Technologies used: SWaV, Convolutional Neural Networks (CNN), Self-Supervised Learning.
ADRENALIN Challenge: Non-Intrusive Load Disaggregation (NILM)
Year: Feb - Apr 2024
Category: Energy Disaggregation, Unsupervised Learning
Competed in the ADRENALIN challenge to predict heating and cooling components in buildings using Non-Intrusive Load Disaggregation (NILM).
Placed 8th worldwide for developing an unsupervised algorithm to accurately disaggregate heating and cooling loads from total power main meter readings.
Technologies used: Unsupervised Learning, NILM, Time-Series Analysis.
Winner HackPlaksha 24' Healthcare Track
Associated with: Plaksha University
Category: Healthcare, Human Computer Interaction
Developed WeCare, a slouching and blinking monitoring desktop app designed to improve user posture and reduce back pain and eye strain by raising alerts when slouching or strain occurs for extended periods.
Technologies used: Strided Transformer weights for body keypoint detection and slouch detection, Electron JS, Python, SQL for desktop app development and data analysis.
Easter Egg: Reinforcement Learning Convergence Simulation
Category: Reinforcement Learning, Temporal Difference Learning
This is not a project but a simulation that visualizes the convergence of a reinforcement learning algorithm, specifically temporal difference learning. The example is inspired by David Silver's RL course slides, and this GIF was generated using Python simulations.
You can tweak the hyperparameters or adjust the state space to observe different convergence behaviors across epochs.