Abstract

Silicon Photonics is a promising technology to develop neuromorphic hardware accelerators. Most optical neural networks rely on wavelength division multiplexing (WDM), which calls for power-hungry calibration to compensate for the non-uniformity fabrication process and thermal variations of microring resonators (MRR). This imposes practical limitations on neuromorphic photonic hardware since only a small number of synaptic connections per neuron can be implemented. As a result, the mapping of neural networks (NN) on a hardware platform requires the pruning of synaptic connections, which drastically affects the accuracy.

In this work, we address these limitations from two directions. First, we proposed a method to map pre-trained NN on an all-optical spiking neural network (SNN). The technique relies on weight partitioning and unrolling to reduce synaptic connections. This method aims to improve hardware utilization while minimizing accuracy loss. The resulting neural networks are mapped on an architecture we propose, allowing us to estimate accuracy and energy consumption. Results show the capability of weight partitioning to implement a realistic NN while attaining a 58\% reduction in energy consumption compared with unrolling.

Second, a synaptic weighting architecture is proposed to implement weighting while reducing the number of required MRRs by half thus simplifying the calibration requirements. The architecture was simulated to demonstrate its capability of performing synaptic weighting. These methods together introduce design directions that can work around constraints of photonic spiking neural network architectures and help reach toward realizing large-scale photonic spiking neural networks.