Recent research has made significant strides in the development of optical neural networks, presenting a sustainable alternative to the energy and resource-intensive models currently in use. By leveraging light propagation through multimode fibers and a minimal number of programmable parameters, researchers have achieved comparable accuracy to traditional digital systems with significantly reduced memory and energy requirements. This innovative approach offers a promising pathway toward energy-efficient and highly efficient artificial intelligence hardware solutions.
A novel architecture for optical neural networks utilizes wavefront shaping to precisely manipulate the travel of ultrashort pulses through multimode fibers, enabling nonlinear optical computation.
Present-day artificial intelligence systems rely on billions of adjustable parameters to accomplish complex objectives. Yet, the vast quantity of these parameters incurs significant expenses. The training and implementation of such extensive models demand considerable memory and processing power, available only in enormous data center facilities, consuming energy on par with the electrical demands of medium-sized cities. In response, researchers are currently reevaluating both the computing infrastructure and the <span class="glossaryLink" aria-describedby="tt" data-cmtooltip="
” data-gt-translate-attributes=”[{"attribute":"data-cmtooltip", "format":"html"}]” tabindex=”0″ role=”link”>machine learning algorithms to ensure the sustainable advancement of artificial intelligence continues at its current rate.
Optical implementation of neural network architectures is a promising avenue because of the low-power implementation of the connections between the units. New research reported in Advanced Photonics combines light propagation inside multimode fibers with a small number of digitally programmable parameters and achieves the same performance on image classification tasks with fully digital systems with more than 100 times more programmable parameters.
This computational framework streamlines the memory requirement and reduces the need for energy-intensive digital processes, while achieving the same level of <span class="glossaryLink" aria-describedby="tt" data-cmtooltip="
” data-gt-translate-attributes=”[{"attribute":"data-cmtooltip", "format":"html"}]” tabindex=”0″ role=”link”>accuracy in a variety of machine learning tasks.
Breakthrough in Nonlinear Optical Computations
The heart of this groundbreaking work, led by Professors Demetri Psaltis and Christophe Moser of EPFL (Swiss Federal Institute of Technology in Lausanne), lies in the precise control of ultrashort pulses within multimode fibers through a technique known as wavefront shaping. This allows for the implementation of nonlinear optical computations with microwatts of average optical power, reaching a crucial step in realizing the potential of optical neural networks.
“In this study, we found out that with a small group of parameters, we can select a specific set of model weights from the weight bank that optics provides and employ it for the aimed computing task. This way, we used naturally occurring phenomena as a computing hardware without going into the trouble of manufacturing and operating a device specialized for this purpose,” states Ilker Oguz, lead co-author of the work.
This result marks a significant stride towards addressing the challenges posed by the escalating demand for larger machine learning models. By harnessing the computational power of light propagation through multimode fibers, the researchers have paved the way for low-energy, highly efficient hardware solutions in artificial intelligence.
As showcased in the reported nonlinear optics experiment, this computational framework can also be put to use for efficiently programming different high-dimensional, nonlinear phenomena for performing machine learning tasks and can offer a transformative solution to the resource-intensive nature of current AI models.
Reference: “Programming nonlinear propagation for efficient optical learning machines” by Ilker Oguz, Jih-Liang Hsieh, Niyazi Ulas Dinc, Uğur Teğin, Mustafa Yildirim, Carlo Gigli, Christophe Moser and Demetri Psaltis, 25 January 2024, Advanced Photonics.
DOI: 10.1117/1.AP.6.1.016002