Optical computing offers a promising solution for achieving high-performance neural networks without relying heavily on electronic computing power. This work introduces a novel framework for implementing programmable neural networks. The framework enables both linear and nonlinear transformations at low optical power by leveraging multiple scattering and data repetition to achieve high-order nonlinearities. By combining linear optics and structural nonlinearity, it offers scalability and programmability for optical computing, particularly in artificial intelligence applications. The framework's ability to synthesize a learnable linear and nonlinear data transform bridges the gap between optical and digital neural networks.
As neural networks (NNs) become more capable, their computational resource requirements also increase exponentially. Optical systems can provide alternatives with higher parallelizability and lower energy consumption. However, the conventional training method, error backpropagation, is challenging to implement with these analog systems since it requires the characterization of the hardware. In contrast, the Forward-Forward Algorithm defines a local loss function for each layer and trains them sequentially without tracking the error gradient between different layers. In this study, we experimentally demonstrate the suitability of this approach for optical NNs by utilizing the multimode nonlinear propagation inside an optical fiber as a building block of the NN. Compared to the all-digital implementation, the optical NN achieves significantly higher classification accuracy while utilizing the optical system only one epoch per layer.
As artificial intelligence (AI) use grows, the demand for computing resources has increased dramatically. Server costs are significant and highlight the need for more energy-efficient and cost-effective computing platforms for AI applications and future growth. One potential solution is optical hardware. Optical computing hardware has several advantages, such as high bandwidth parallelism and energy efficiency. However, one major limitation is the implementation of nonlinear calculations in the optical domain.
We will discuss an approach that achieves the equivalent of optical nonlinearity vastly more effectively than current approaches. The essence of the technique relies on multiple linear scattering off data encoded onto a spatial light modulator (SLM) that uses low optical power to effectively synthesise a nonlinear operation. By exploiting this relationship, arbitrary nonlinear transformations are programmed digitally, and light effectively performs an all-optical computation without requiring electronic switching or high peak power to achieve non-linearity.
The ever-increasing demand for training and inferring with larger machine-learning models requires more efficient hardware solutions due to limitations such as power dissipation and scalability. Optics is a promising contender for providing lower power computation, since light propagation through a nonabsorbing medium is a lossless operation. However, to carry out useful and efficient computations with light, generating and controlling nonlinearity optically is a necessity that is still elusive. Multimode fibers (MMFs) have been shown that they can provide nonlinear effects with microwatts of average power while maintaining parallelism and low loss. We propose an optical neural network architecture that performs nonlinear optical computation by controlling the propagation of ultrashort pulses in MMF by wavefront shaping. With a surrogate model, optimal sets of parameters are found to program this optical computer for different tasks with minimal utilization of an electronic computer. We show a remarkable decrease of 97% in the number of model parameters, which leads to an overall 99% digital operation reduction compared to an equivalently performing digital neural network. We further demonstrate that a fully optical implementation can also be performed with competitive accuracies.
Utilizing light propagation and optical nonlinearities is one of the strategies to accelerate computational tasks in tandem with electrical circuits in an energy-efficient manner. Computing with multimode optical fibers has been demonstrated to be energy efficient due to the high light confinement and multidimensionality. However, these optical nonlinearities have not been programmed for a specific computational task and thus the performance is not optimal. In this study, we demonstrate that the nonlinear transformation in the fiber can be programmed to obtain improved performances on several different machine learning tasks by shaping the wavefront of the information encoding beam.
Single-mode optical waveguides are one of the fundamental photonic components and the building block for compact multicore bundles. The cross-talk of a waveguide bundle might scramble the information and reduce the resolution. To get the highest possible resolution on a fixed field of view (FOV), we propose to optimize the core to core spacing via assessing the reconstruction accuracy of the output images of the bundle processed by a deep neural network (DNN), where the obtained bundle is 3D printed via direct laser writing. We demonstrate the DNN based optimization scheme and the fabrication of a waveguide bundle of 10-µm core-to-core spacing to image various digit layouts in a (120 µm)^2 FOV.
Advancements in 3D printing, specifically by 2-photon polymerization, enabled the fabrication of high resolution novel optical elements for various applications by engineering the topography of the structures. Recently, the ability to obtain a varying refractive index distribution by 2-photon polymerization also started to gain momentum to design and produce gradient-index (GRIN) micro-optics. Here, we demonstrate micro-scale volume holograms by 2-photon polymerization by tuning the exposure point by point in the 3D volume to obtain the necessary refractive index distribution. 3D printed micro-scale GRIN volume holograms lay an opportunity to open a complimentary dimension in various systems thanks to in-situ fabrication advantage, which can provide complex interconnections and beam shaping in micro-scale.
Obtaining a varying refractive index distribution has always been attracting a high interest in the optics community to produce gradient-index (GRIN) optics. The conventional way to store and process data by GRIN media is through volume holograms, where the recording is done by optical means, which prevents independently accessing each point in the volume. Additive manufacturing, specifically 2-photon polymerization, inherits this ability. Considering the scalability advantage of the 3D implementation of computation architectures and the power-speed advantage of optics, there lie many opportunities for additively manufactured GRIN optics performing complex tasks. Independent access to each voxel in fabrication volume opens the way for digital optimization techniques to design GRIN optics since each calculated voxel can be translated into fabrication. In this work, Learning Tomography (LT), which is a nonlinear optimization algorithm originally developed for optical diffraction tomography, is used as the optimization framework to calculate necessary refractive index distribution to perform computation tasks such as matrix multiplication. Here, instead of imaging an object in optical diffraction tomography, we calculate the 3D GRIN element that performs the desired task as defined by its input-output relation. This input-output relation can be chosen such that a computational functionality is satisfied. We report functional robust GRIN elements where the refractive index dynamic range (<0.005) is comparable to the dynamic range of conventional holography materials. We present the digital optimization methodology with details on the beam propagation method as the forward model and the corresponding error reduction scheme for the desired input-output mapping along with the experimental verification of the approach along with the details of the fabrication process by additive manufacturing.
Volume holography has been widely investigated for information storage and other applications. An increase in the number of multiplexed holograms leads to dynamic range loss when they are optically recorded. Computer generated holograms can achieve better performance if constructed voxel-by-voxel or as a multilayer structure. Advancements in 3D printing enabled the fabrication of multilayered diffractive elements in the micro-scale. To obtain an accurate design, we deploy the Learning Tomography (LT) method, which is an optimization algorithm for computationally imaging 3D distribution of the refractive index. Here, instead of imaging an object, we define a 3D structure that achieves a desired functionality.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.