Modeling Silicon-Photonic Neural Networks under Uncertainties

Abstract

Silicon-photonic neural networks (SPNNs) offer substantial improvements in computing speed and energy efficiency compared to their digital electronic counterparts. However, the energy efficiency and accuracy of SPNNs are highly impacted by uncertainties that arise from fabrication-process and thermal variations. In this paper, we present the first comprehensive and hierarchical study on the impact of random uncertainties on the classification accuracy of a Mach-Zehnder Interferometer (MZI)based SPNN. We show that such impact can vary based on both the location and characteristics (e.g., tuned phase angles) of a non-ideal silicon-photonic device. Simulation results show that in an SPNN with two hidden layers and 1374 tunable-thermalphase shifters, random uncertainties even in mature fabrication processes can lead to a catastrophic 70% accuracy loss.

LoCI: An Analysis of the Impact of Optical Loss and Crosstalk Noise in Integrated Silicon-Photonic Neural Networks

“…While several high-performance AI accelerators based on coherent SP-NNs have been recently proposed [2,10,13], [14] showed that the inferencing accuracy of SP-NNs can drop by up to 70% due to fabrication-process variations and thermal crosstalk. In addition to these variations, the work in [10] explored the impact of optical loss non-uniformity among MZIs and showed SP-NN performance degradation.…”

“…1(a)), where the number of MZIs depends on the OIU architecture [6]. Note that 𝜃 and 𝜙 in each MZI, where 𝜃 determines the state and hence optical loss and crosstalk noise introduced in each MZI, depend on the weight parameters and can be determined using SP-NN training algorithms [14]. The output of the OIU is connected to an optical-gain unit (OGU) that includes semiconductor optical amplifiers (SOAs) [7].…”

“…To analyze the system-level impact of optical loss and crosstalk, we consider a case study of an SP-NN with two hidden layers (𝑀 = 3) of 16 neurons each (𝑁 = 16), trained on the MNIST handwritten digit classification task. Each image in the MNIST dataset is converted to a complex feature vector of length 16 using the fast Fourier transform [14]. The nominal test accuracy is 93.86%.…”

Optimizing Coherent Integrated Photonic Neural Networks under Random Uncertainties

“…As a result, any deviations in such adjusted phase settings (i.e., weights) will lead to network inferencing-accuracy losses. Unfortunately, the underlying SiPh devices in IPNNs are sensitive to inevitable fabrication-process variations (FPVs) and on-chip thermal crosstalk [2], both of which are a significant source of phase errors in IPNNs, and can cause up to 70% loss in IPNN inferencing-accuracy [3]. Leveraging the non-uniqueness of SVD under reflections, we propose a novel optimization method to improve power efficiency and robustness in SiPh-based coherent IPNNs under random uncertainties-stemming from FPVs and thermal crosstalk-without affecting the inferencing-accuracy.…”

“…To verify this at the system-level, we consider a four-layer IPNN and observe its performance on the MNIST dataset. Each real-valued image in the MNIST dataset is converted to a complex feature vector of length 16 using a method based on fast Fourier transform [3]. A fully-connected network with two hidden layers of 16complex valued neurons is implemented using the Clements design [1].…”