An efficient optical-based binary neural network hardware accelerator for harsh environments
Belal Jahannia, Jiachi Ye, Salem Altaleb, Chandraman Patil, Elham Heidari, Hamed Dalir · 2024
Binarized neural networks offer substantial reductions in memory and computational requirements compared to full precision networks. However, conventional CMOS-based hardware implementations still face challenges with resilience for deployment in harsh environments like space. This paper proposes an optical XOR-based accelerator for binarized neural networks to enable low power and resilient operation. The optical logic gates rely on wavelength-specific intensity propagation rather than absolute intensity levels. This provides inherent robustness against fabrication process variations and high energy particle strikes. Simulations of an optical hardware prototype for XNOR-Net show the accelerator achieves 1.2 μs latency and 3.2 mW power. The binarized network maintained 2-4% accuracy degradation compared to the full precision baseline on MNIST and CIFAR-10. The proposed optical accelerator enables efficient and resilient deployment of binarized neural networks for harsh environment applications like spacecraft and satellites.