Free-space optical neural network based on thermal atomic nonlinearity

Abstract

As artificial neural networks (ANNs) continue to make strides in wide-ranging and diverse fields of technology, the search for more efficient hardware implementations beyond conventional electronics is gaining traction. In particular, optical implementations potentially offer extraordinary gains in terms of speed and reduced energy consumption due to intrinsic parallelism of free-space optics. At the same time, a physical nonlinearity -a crucial ingredient of an ANN -is not easy to realize in free-space optics, which restricts the potential of this platform. This problem is further exacerbated by the need to perform the nonlinear activation also in parallel for each data point to preserve the benefit of linear free-space optics. Here, we present a free-space optical ANN with diffraction-based linear weight summation and nonlinear activation enabled by the saturable absorption of thermal atoms. We demonstrate, via both simulation and experiment, image classification of handwritten digits using only a single layer and observed 6% improvement in classification accuracy due to the optical nonlinearity compared to a linear model. Our platform preserves the massive parallelism of free-space optics even with physical nonlinearity, and thus opens the way for novel designs and wider deployment of optical ANNs.

High-Speed Multi-Layer Convolutional Neural Network Based on Free-Space Optics

“…One of the most challenging part of implementing the ONNs is realizing physical optical nonlinearity [25]. Saturable absorption nonlinearity is an all-optical nonlinearity which can be performed in free-space optics by passing light through an atomic vapor.…”

“…Saturable absorption nonlinearity is an all-optical nonlinearity which can be performed in free-space optics by passing light through an atomic vapor. An atomic vapor cell is a glass cell containing a specific gas, which represents specific absorption spectrum, and it provides nonlinear relation between its input and output propagated lights [25]. ( 12) and ( 13) represent the mathematical model of saturable absorber (SA) [32] and its derivative, presuming a real-valued E-field (otherwise the square term should be the square of the absolute value, i.e., |E| 2 ):…”

“…Recently, some new interests have been activated in developing ONN, and new ONNs have been proposed for implementing photonic multilayer perceptrons (MLPs) [10], [16]- [18] as well as photonic CNNs [6], [7], [19]- [21]. It should be noted that some implementation is based on integrated setup [22], [23], while the others are based on free-space one [19], [20], [24], [25]. Although the integrated structures offer lower power consumption and lower area in comparison with the free-space structures, they suffer from reconfigurability and scalability problems, as well as fabrication challenges.…”

Deep learning in photonics: introduction

“…Nanostructured photonic devices can exploit subwavelength linear and nonlinear scatterers to realize complex input-output mapping far beyond the capabilities of traditional nanophotonic devices. Theoretical and experimental exploration works in this feature issue have discussed neuromorphic node based on quantum dot laser [13], end-to-end optical backpropagation for training neural networks [14], free-space optical neural network based on thermal atomic nonlinearity [15], deep-learning empowered dynamic wavefront shaping in nonstationary scattering media [16], delay weight plasticity based on supervised learning in photonic spiking neural networks [17], and all-optical neuromorphic binary convolution with spiking VCSEL neurons [18].…”

Co-Design of Free-Space Metasurface Optical Neuromorphic Classifiers for High Performance

“…While improvements are possible using neural network algorithms, the approach still requires significant resources. Recently, an approach to circumvent both limitations has been proposed based on passive metamaterials that diffract free-space optical fields. − The free-space propagation allows information processing and transfer at the speed of light until it is captured by a few low-power and fast photodetector pixels. Thus, overall speed and energy efficiency can in principle be significantly improved compared with the conventional optoelectronic approach.…”