Programming light propagation creates highly efficient neural networks

Current AI models use billions of parameters that can be trained to accomplish difficult tasks. However, this large number of parameters comes at a high cost. Training and deploying these enormous models requires immense memory space and computing capabilities that can only be provided by hangar-sized data centers in processes that consume energy equivalent to the electricity needs of cities medium sized.

The research community is currently working to rethink both the associated computing hardware and machine learning algorithms in order to sustainably maintain the development of artificial intelligence at its current pace. The optical implementation of neural network architectures is a promising avenue due to the low consumption of connections between 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 computing framework streamlines memory requirements and reduces the need for energy-intensive digital processes, while achieving the same level of accuracy in a variety of machine learning tasks.

The heart of this work, led by Professors Demetri Psaltis and Christophe Moser of EPFL (Ecole Polytechnique Fédérale de Lausanne), lies in the precise control of ultrashort pulses within multimode fibers using a technique known as wavefront shaping. This enables the implementation of nonlinear optical calculations with microwatts of average optical power, taking a crucial step in realizing the potential of optical neural networks.

“In this study, we found that with a small group of parameters, we can select a specific set of model weights from the weight bank provided by the optics and use it for the targeted computational task. In this way, we used natural phenomena as hardware without going to the trouble of manufacturing and operating a specialized device for this purpose,” says Ilker Oguz, co-lead author of the work.

This result marks an important step toward solving the challenges posed by the growing demand for larger machine learning models. By harnessing the computing power of light propagation through multimode fibers, researchers have paved the way for low-power, highly efficient artificial intelligence hardware solutions.

As shown in the reported nonlinear optics experiment, this computational framework can also be used to efficiently program different high-dimensional nonlinear phenomena to perform machine learning tasks and can offer a transformative solution to the greedy nature in resources of current AI models.