Symmetric silicon microring resonator optical crossbar array for accelerated inference and training in deep learning

TL;DR

This work tackles the challenge of enabling on-chip backpropagation for photonic neural accelerators by introducing a symmetric silicon microring resonator (MRR) crossbar. The authors design and demonstrate a $4\times4$ MRR crossbar on a silicon-on-insulator platform, achieving $93.3\%$ Iris classification accuracy during inference, and show, via simulated on-chip backpropagation, a training accuracy of $91.1\%$ on the same task. They further simulate a $9\times9$ MRR crossbar to perform CNN-like convolution operations, achieving $93.4\%$ accuracy on MNIST after training with ADAM, illustrating the potential of photonic tensor cores for both accelerated inference and on-chip learning. The results highlight a path toward compact, energy-efficient photonic accelerators capable of scaling to more complex networks, with improvements anticipated from faster, low-power phase shifters and larger, multi-wavelength implementations.

Abstract

Photonic integrated circuits are emerging as a promising platform for accelerating matrix multiplications in deep learning, leveraging the inherent parallel nature of light. Although various schemes have been proposed and demonstrated to realize such photonic matrix accelerators, the in-situ training of artificial neural networks using photonic accelerators remains challenging due to the difficulty of direct on-chip backpropagation on a photonic chip. In this work, we propose a silicon microring resonator (MRR) optical crossbar array with a symmetric structure that allows for simple on-chip backpropagation, potentially enabling the acceleration of both the inference and training phases of deep learning. We demonstrate a $4 \times 4$ circuit on a Si-on-insulator (SOI) platform and use it to perform inference tasks of a simple neural network for classifying Iris flowers, achieving a classification accuracy of 93.3%. Subsequently, we train the neural network using simulated on-chip backpropagation and achieve an accuracy of 91.1% in the same inference task after training. Furthermore, we simulate a convolutional neural network (CNN) for handwritten digit recognition, using a $9 \times 9$ MRR crossbar array to perform the convolution operations. This work contributes to the realization of compact and energy-efficient photonic accelerators for deep learning.

Figures (10)

• Figure 1: Proposed optical crossbar array. The matrix and vector are generated by MRRs and MZIs, respectively. Multiple wavelengths are injected into 4 input ports simultaneously. The MRRs are tuned to align with different wavelengths, and the associated matrix element is represented by the transmittance of optical power at the drop port. (a) By injecting a forward signal $\boldsymbol x$, which represents the output signal from the previous layer in an ANN, the crossbar array performs the multiplication between $\mathrm{\bf W}$ and $\boldsymbol x$. (b) By injecting a backward signal $\boldsymbol \sigma$, which represents the error signal backpropagated from the next layer in an ANN, the crossbar array performs the multiplication between $\mathrm{\bf W^\top}$ (the transpose of $\mathrm{\bf W}$) and $\boldsymbol \sigma$.
• Figure 2: Microscope images of a $4 \times 4$ MRR crossbar array fabricated on a SOI platform. (a) The entire circuit consists of MZIs, MRRs, and TE-pass filters. The consumed chip area is 3.7 $\times$ 2.4 ^2. (b) Enlarged view of one MZI. Only one input port and one output port are used. The other two ports are terminated with inverse waveguide tapers. (c) Enlarged view of one MRR. The radii of all MRRs are 20 .
• Figure 3: Experimental setup. Four CW lights at different wavelengths are generated by a 4-channel tunable laser and combined into a single optical fiber by inversely using two stages of $1 \times 2$ optical splitters. The MEMS optical switch directs the combined light to the ports for either the forward or backward signal. The chip is wire-bonded for external electrical control and packaged with a fiber array for stable fiber coupling.
• Figure 4: Characterizations of MZIs and MRRs. (a) Characterization result of the MZI in the In 1 port as a function of heater power. The MZI exhibits a high extinction ratio of 51 dB. (b) Transmission spectra measured at the Out 1-4 ports when sweeping the wavelength of light injected into the In 1 port. No electric power is applied to the phase shifters for MRRs. The resonant wavelengths of the four MRRs slightly differ due to fabrication non-uniformity. (c) Illustration of characterizing the difference between the forward and backward paths of each MRR. (d) Optical power measured at the output ports for forward and backward signals of one MRR. The two directions exhibit almost the same characteristics.
• Figure 5: Experimental implementations of various matrices for forward and backward signals. Each matrix element is measured by setting one MZI into the maximum-transmittance state and the others into the minimum-transmittance state. The matrices measured from the forward and backward directions are transposed to each other. Error signals in these matrices are suppressed to the level of approximately -15 dB.