Bio-Adaptive Music Engine (BAME)
A Bio-Adaptive Music Engine that uses Deep Reinforcement Learning (SAC) and a Transformer-based World Model to autonomously select music from a high-dimensional audio embedding space (MERT), optimizing for real-time regulation of a user's physiological state.
ReMERT: Enhancing Learning Efficiency in Deep Q-Learning
PyTorch implementation of DQN enhanced with ReMERT (Replay Memory with End-Related Transitions). Replaces uniform experience replay with a prioritized sampling strategy based on inverse distance to terminal states, significantly improving sample efficiency and convergence speed on the CartPole-v1 environment.
Licence Plate Image Recognizer
Full pipeline for license plate detection and recognition, using YOLOv5 for detection and two recognition models: CNN + LSTM + CTC baseline and transformer-based PDLPR model.
OCR Post-Correction and Evaluation Pipeline
Fine-tuned LLMs with Custom Evaluation Framework.
Proteins-GNN-Classifier
Graph Neural Network for Proteins Classification with different levels and types of label noise.
Water Body Segmentation
AER U‑Net: Attention‑Enhanced Multi‑Scale Residual U‑Net for Water Body Segmentation.