Lighting Up AI: A Sustainable Future
The current trend in artificial intelligence involves using models with billions of trainable parameters, requiring substantial computational resources and energy consumption. Researchers are exploring ways to make AI development more sustainable by rethinking computing hardware and algorithms. One promising approach is the optical implementation of neural network architectures, leveraging the low power consumption of connections between units.
A recent study in Advanced Photonics, led by Professors Demetri Psaltis and Christophe Moser at EPFL, introduces a novel computational framework. This framework combines light propagation inside multimode fibers with a small number of digitally programmable parameters, achieving comparable performance in image classification tasks to fully digital systems with over 100 times more parameters. The technique reduces memory requirements and lessens the need for energy-intensive digital processes, maintaining accuracy across various machine learning tasks.
The core of the research lies in precise control of ultrashort pulses within multimode fibers using wavefront shaping. This enables the implementation of nonlinear optical computations with microwatts of average optical power, a crucial advancement for optical neural networks.
Lead co-author Ilker Oguz explains that the study found that a small set of parameters can be used to select specific model weights from the optical weight bank, utilizing naturally occurring phenomena as computing hardware without the need for specialized devices.
This breakthrough addresses challenges associated with the increasing demand for larger machine learning models. By tapping into the computational power of light propagation through multimode fibers, the researchers have opened the door to low-energy, highly efficient hardware solutions for artificial intelligence. The demonstrated computational framework can efficiently program various high-dimensional, nonlinear phenomena for machine learning tasks, offering a transformative solution to the resource-intensive nature of current AI models.
Смотрите видео Lighting Up AI: A Sustainable Future онлайн без регистрации, длительностью часов минут секунд в хорошем качестве. Это видео добавил пользователь The Tech Savvy Society 31 Январь 2024, не забудьте поделиться им ссылкой с друзьями и знакомыми, на нашем сайте его посмотрели 29 раз и оно понравилось 0 людям.