Research Work
QUESC: Environmental Sound classification Using Quantum Quantized Networks
Description: This research develops a Quantized Hybrid Quantum-Classical Neural Network (QQNN) approach for Environmental Sound Classification that combines pre-trained classical layers with quantum circuits, achieving comparable accuracy to traditional methods while demonstrating quantum computing’s viability in audio classification tasks.
Key Highlights:
- Introduces a Quantized Hybrid Quantum-Classical Neural Network (QQNN) for environmental sound classification, combining classical and quantum layers to leverage quantum computing’s advantages.
- Demonstrates a 21.17% performance improvement in training compared to traditional models, showing promise for environmental sound classification.
- Employs quantization to reduce memory usage significantly, enabling compatibility with edge devices and making the model efficient for real-time applications.
- Suggests future applications in quantum transfer learning and a potential transition to fully quantum-based models for more efficient ESC tasks.
Drones for Post-Flood Rescue Missions
Description: This research proposes the use of autonomous drones equipped with advanced computer vision for post-flood rescue missions, enabling efficient detection and localization of stranded individuals. By integrating high-resolution cameras, GPS, and machine learning algorithms, the system aims to enhance search capabilities and improve response times during critical post-disaster periods.
Key Highlights:
- Introduces an autonomous drone system for post-flood rescue missions, featuring computer vision technology to improve detection accuracy and operational efficiency in disaster zones.
- Demonstrates a modular design combining Pixhawk flight controllers, RGB cameras, and machine learning models (YOLOv8 and DETR-ResNet50) to accurately identify and locate stranded persons.
- Reduces rescue operation time and resource allocation by employing geofencing and pre-flood area mapping, prioritizing high-risk zones for targeted searches.
Evolutionary Algorithms for Neural Network Training
Description: This publication proposes using Evolutionary Algorithms in combination with Backpropagation for training Neural Networks. The study evaluates the effectiveness, efficiency, and impact on model accuracy, aiming to identify the optimal approach for neural network training.
Key Highlights:
- Evaluates various Evolutionary Algorithms like Genetic Algorithms, Differential Evolution etc. and their variants.
- Proposes A novel hybrid training mechanism, enabling an adaptive choice between EAs and Backpropagation based on real-time training metrics.
- Improves the model’s ability to generalize to unseen data through Evolutionary Algorithms, showing a reduction in overfitting.
- Includes an automated process for tuning hyperparameters, such as mutation rates in EAs and learning rates in Backpropagation.
- Explores applications of this hybrid approach within complex architectures, such as Convolutional Neural Networks (CNNs).