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.
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.
As light propagates through a multimode fiber, the optical modes exchange power, create new optical frequencies via a complex spatiotemporal non-linear transformation that occurs at relatively low optical powers because of the light confinement and the long interaction length that is possible in a fiber. In this talk, we will review exciting new developments in this field including from our research group.
In particular, we show recent results of programming the nonlinear interaction in MMFs for machine learning applications. Several different databases were used to train a system consisting of a multimode fiber of different length/core size of a MMF and a simple, single layer digital network. In several recognition tasks, the classification accuracy that was obtained was comparable to deep, digitally implemented networks. The energy requirement for training and reading the optical system was orders of magnitude less than the digital counterpart.
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.
An optical computing framework based on spatiotemporal nonlinear effects of multimode fibers is presented. Experimentally, a powerful computation engine can be realized using linear and nonlinear interactions of spatial fiber modes. With the present optical scheme, we demonstrated excellent performance on a variety of classification and regression tasks. Our studies showed that spatiotemporal fiber nonlinearities are as effective as digital neural network structures in challenging computational tasks. Featuring better energy efficiency and easy scalability, our method provides a new approach to optical computation.
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.