The diamond mesh, a phase-error- and loss-tolerant field-programmable MZI-based optical processor for optical neural networks

Abstract

This paper presents the performance analysis of a phase error- and loss-tolerant multiport field-programmable MZI-based structure for optical neural networks (ONNs). Compared to the triangular (Reck) mesh, our proposed diamond mesh makes use of a larger number of MZIs, leading to a symmetric topology and adding additional degrees of freedom for the weight matrix optimization in the backpropagation process. Furthermore, the additional MZIs enable the diamond mesh to optimally eliminate the excess light intensity that degrades the performance of the ONNs through the tapered out waveguides. Our results show that the diamond topology is more robust to the inevitable imperfections in practice, i.e., insertion loss of the constituent MZIs and the phase errors. This robustness allows for better classification accuracy in the presence of experimental imperfections. The practical performance and the scalability of the two structures implementing different sizes of optical neural networks are analytically compared. The obtained results confirm that the diamond mesh is more error- and loss-tolerant in classifying the data samples in different sizes of ONNs.

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

“…Several architectures have been proposed to enable MZI-based linear multipliers (i.e., OIU in Fig. 1(a)) for deep neural networks [6,9,10]. A fully connected layer 𝐿 𝑚 with 𝑛 𝑚 neurons performs linear multiplication between an input vector and a weight matrix (𝑊 ) followed by a non-linear activation (𝑓 ).…”

“…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. As the size and complexity of emerging SP-NNs increase to handle more complex tasks, the total insertion loss in the network increases as well.…”

Characterization and Optimization of Integrated Silicon-Photonic Neural Networks under Fabrication-Process Variations

“…Several approaches have been proposed to derive the architecture of MZI arrays to perform MAC operations in the optical domain [8]- [10]. Out of these, the Clements design, due its symmetric nature, has the lowest optical loss and footprint.…”

Photonic Integrated Reconfigurable Linear Processors as Neural Network Accelerators

“…Alternatively, in the "Clements" architecture, the same number of MZIs are arranged in a rectangular shape [35], as exemplified in Figure 2b. With the "Clements" architecture, optical path lengths are balanced and the processor depth is reduced to N. Recently, a diamond architecture has been proposed as an extended version of the "Reck" topology [36], obtained by adding (N − 1)(N − 2)/2 MZIs to the triangular one. This architecture, depicted in Figure 2c, requires (N − 1) 2 MZIs, and has the same processor depth as the "Reck", 2N − 3.…”

“…Both the "Clements" and the diamond architectures are more robust to imperfections, noise, and loss with respect to the "Reck" one, translating into higher accuracies when used in neuromorphic applications [35,36]. However, the diamond mesh requires more MZIs, thus resulting in less power and footprint, and a more complex calibration.…”