Hardware-Algorithm Co-Optimization of Early-Exit Neural Networks for Multi-Core Edge Accelerators
The research optimizes Early-Exiting Neural Networks (EENN) for edge accelerators by treating deployment as a multi-objective problem balancing accuracy and energy-latency costs. By using a genetic algorithm guided by hardware modeling, the framework found EENN configurations achieving over 50% reduction in energy-latency product compared to static models under 8-bit quantization, demonstrating efficient hardware co-optimization.