Towards On-Chip Optical FFTs for Convolutional Neural Networks

Abstract

Convolutional neural networks have become an essential element of spatial deep learning systems. In the prevailing architecture, the convolution operation is performed with Fast Fourier Transforms (FFT) electronically in GPUs. The parallelism of GPUs provides an efficiency over CPUs, however both approaches being electronic are bound by the speed and power limits of the interconnect delay inside the circuits. Here we present a silicon photonics based architecture for convolutional neural networks that harnesses the phase property of light to perform FFTs efficiently. Our all-optical FFT is based on nested Mach-Zender Interferometers, directional couplers, and phase shifters, with backend electro-optic modulators for sampling. The FFT delay depends only on the propagation delay of the optical signal through the silicon photonics structures. Designing and analyzing the performance of a convolutional neural network deployed with our on-chip optical FFT, we find dramatic improvements by up to 10 4 when compared to state-of-the-art GPUs when exploring a compounded figure-of-merit given by power per convolution over area. At a high level, this performance is enabled by mapping the desired mathematical function, an FFT, synergistically onto hardware, in this case optical delay interferometers.

Photonic convolutional neural networks using integrated diffractive optics

“…Such a direct implementation of N × N DFT requires N (N −1) 2 MZIs, each consisting of twice that number of phase shifters and beam splitters. A more efficient design using the Cooley-Tukey FFT algorithm reduces the number of MZIs needed to N log 2 (N ) 2 [31], [32]. However, the waveguide crossings needed grows as a triangular number T N 2 −1 , which can introduce a significant insertion loss.…”

Photonic convolutional neural networks using integrated diffractive optics

“…[31], [32]. However, the waveguide crossings needed grows as a triangular number T N 2 −1 , which can introduce a significant insertion loss.…”

Frequency Domain Interleaving for Dense WDM Passive Optical Network

“…The demultiplexing operation is introduced at the ONU side and is processed by the use of an alloptical passive element (Mach-Zehnder interferometer -MZI). For practical purposes, this processing 198 can be implemented with silicon-on-insulator (SOI) technology in photonic integrated circuits (PICs) [6][7]. In its most basic form, the technology allows to implement passive optical components, such as splitters, filters, (de)multiplexers, polarization handling components, interferometers, resonators and their combination with coupling structures to optical fibers or to free space optical elements [7].…”

“…In this case, all MZIs that are not part of the optical path to the corresponding output port can be removed, leaving only one MZI per stage [16]. Another possibility for practical implementation relies on, as aforementioned, implementation with silicon-on-insulator technology, obtaining much more compact structures [6][7]. We used a standard single mode fiber (SMF) model with attenuation factor α = 0.…”