Highly efficient convolution computing architecture based on silicon photonic Fano resonance devices

Abstract:

Convolutional neural networks (CNNs) require a lot of multiplication and addition operations completed by traditional electrical multipliers, leading to high power consumption and limited speed. Here, a silicon waveguide-based wavelength division multiplexing (WDM) architecture for CNN is optimized with high energy efficiency Fano resonator. Coupling of T-waveguide and micro-ring resonator generates Fano resonance with small half-width, which can significantly reduce the modulator power consumption. Insulator dataset from state grid is used to test Fano resonance modulator-based CNNs. The results show that accuracy for insulator defect recognition reaches 99.27% with much lower power consumption. Obviously, our optimized photonic integration architecture for CNNs has broad potential for the artificial intelligence hardware platform.