Compact Neural Architectures for Edge-Based Facial Recognition

This STTR Phase I project focuses on developing Edge-Optimized Compact Neural Networks (EOCNN) for real-time facial recognition on edge devices. We'll utilize spectral pooling and group convolution for effective feature extraction, and employ dynamic quantization and pruning to streamline network size without sacrificing precision. Reinforcement learning refines our Neural Architecture Search (NAS) for optimal model accuracy. Advanced features, including Sigmoid-based attention, spatial transformer layers, and Model-Agnostic Meta-Learning (MAML), enhance focus, spatial adaptation, and quick task-specific adjustments. With a device profiling mechanism and an edge-centric Lovasz function, the EOCNN promises adaptability and precise predictions, setting a new standard in edge-based facial recognition.