We propose an experimental demonstration of a silicon-based neuromorphic computing scheme with optical interconnects. The device consists of 4-channel input grating coupler arrays to guide 1550 nm light through the waveguide, tunable Mach-Zehnder Interferometer (MZI) mesh for matrix-vector multiplication, micro-ring resonator (MRR)-based MZI to implement nonlinear activation function, and 4-channel radiator array for free-space radiation. Depending on the different input information (i.e., image, voice, text, etc.), the radiated beam is focused in different directions to perform the classification task. Our proposed on-chip optical computing scheme can pave the way for future AIs, providing a small footprint, high-speed, cost-effective, and power-efficient.
|