2024-07-26 06:00:11
The future of artificial intelligence (AI) could be light.
A promising new method, developed by researchers at the Max Planck Institute for the Science of Light, proposes to use an optical system to implement neural networks. This innovative approach could make machine learning faster and much less energy-intensive.
Currently, machine learning and AI technologies are expanding rapidly, from vision tasks to text generation, as illustrated by ChatGPT. However, this resource-intensive growth has led to increasingly complex neural networks, sometimes containing billions of parameters, which is energy unsustainable.
To address this issue, the field of neuromorphic computing seeks to replace digital neural networks with physical networks that can perform the same mathematical operations more quickly and efficiently. Optical and photonic platforms are particularly promising due to their low energy consumption and their vitesse calculation limited only by the speed of light.
However, two major challenges remain: the need for high laser powers to perform complex mathematical calculations and the lack of an efficient learning method for these physical networks. The method proposed by Clara Wanjura and Florian Marquardt overcomes these obstacles by modifying the transmission of light.
This approach avoids the complicated physical interactions required for mathematical functions, thus reducing the need for high light power. In addition, training and inference (exploitation) of the network becomes simpler, simply by sending light through the system and observing the transmitted light.
Fully nonlinear neuromorphic system with linear wave propagation.
Credit: Nature Physics (2024). DOI: 10.1038/s41567-024-02534-9
The researchers’ simulations demonstrated that this method can be used for image classification tasks with an accuracy equivalent to that of digital networks. Next steps include experimental collaborations to explore the implementation of this method on different platforms, opening new possibilities for neuromorphic devices.
1722006192
#ChatGPT #AIs #incredibly #faster #efficient #light #technology
