Efficient Software Implementation of Deep Learning Architectures on Embedded Systems

This scientific talk discusses methods for optimizing deep learning architectures on embedded systems. It highlights key challenges, such as limited processing power, memory constraints, and real-time performance requirements. Model compression techniques, including quantization, pruning, knowledge distillation, and weight sharing, are explored to reduce memory usage and computational complexity. Hardware-software co-design is emphasized, leveraging specialised accelerators like NPUs, GPUs, and FPGAs to improve efficiency. Additionally, software optimization techniques, demonstrated through a radar-based hand gesture recognition project, showcase how deep learning can be effectively deployed on edge devices while balancing accuracy, performance, and resource constraints.