FOCUS
WAVEFRONT control
WAVEFRONT SHAPING WITH STRUCTURAL NONLINEARITY FOR OPTICAL RECURRENT NEURAL NETWORKS
///////////////////////////////////////////////////////////////////////////////////////////////////
Leandro RISSO VENANCIO 1, 2, Jonathan DONG 3, Jacopo BERTOLOTTI 2, Mickael MOUNAIX 1, *
1
School of EECS, The University of Queensland, Brisbane QLD 4072, Australia
2
University of Exeter- Physics Building, Streatham Campus, Stocker Rd, Exeter EX4 4QL, UK
3
Biomedical Imaging Group, École Polytechnique Fédérale de Lausanne, Station 17, 1015 Lausanne, Switzerland * m. mounaix @ uq. edu. au
https:// doi. org / 10.1051 / photon / 202614034
With the exponential growth in demand for computing power, photonics offers a sustainable and competitive alternative to conventional electronic computing. In this article, we present how wavefront shaping, by means of a spatial light modulator, provides a versatile platform for optical computing. Using only linear optical components, we introduce an approach that exploits structural nonlinearity to enhance the forecasting of a chaotic time series.
WAVEFRONT SHAPING FOR OPTICAL COMPUTING The growing demand for computing power, driven in particular by machine learning and artificial intelligence, is putting increasing pressure on conventional electronic hardware, in terms of both computation capacity and energy consumption. Optical computing offers an attractive alternative, exploiting the high speed, high bandwidth, low propagation losses and inherent parallelism of light to perform computations directly in the optical domain [ 1 ].
Wavefront shaping, initially developed for imaging through scattering media, provides a natural platform for optical computing. Information can be encoded using a spatial light modulator( SLM) onto the many spatial degrees of freedom of an optical field. Complex optical systems, such as multiple scattering materials or multimode optical fibers, are particularly attractive for this purpose, as their strong mode mixing in all dimensions( space, polarization and wavelength) can implement large transformations in parallel with a relatively small physical footprint.
Optical computing performs the heavy processing of input data into a representation from which the desired information can be extracted with simple linear models. While earlier works demonstrated the optical implementation of feedforward architectures [ 2 ], reservoir computing provides an ideal framework for optical recurrent neural networks( RNNs). Rather than training the RNN internal connections, the input time series is projected into a high-dimensional space by a fixed nonlinear transformation, with only the last output layer trained [ 3 ]. These
34 www. photoniques. com I Photoniques 140