by Researchers at the Swiss Federal Institute of Technology in Lausanne (EPFL) have developed a groundbreaking method to create highly efficient neural networks using light propagation. The new approach, reported in Advanced Photonics, significantly reduces the number of parameters needed while maintaining the same level of performance as traditional digital systems.
Artificial intelligence models currently rely on billions of trainable parameters to tackle complex tasks. However, these large models require extensive memory space and computing power, which can only be provided by enormous data centers consuming vast amounts of energy. To address these issues, the research community is exploring alternative computing hardware and machine learning algorithms.
The EPFL researchers’ novel technique combines light propagation inside multimode fibers with a small number of programmable parameters, achieving equivalent performance in image classification tasks compared to fully digital systems with over 100 times as many parameters. This breakthrough computational framework not only streamlines memory requirements but also reduces the energy-intensive processes associated with digital systems.
At the core of this innovation is the precise control of ultrashort pulses within multimode fibers using a technique called wavefront shaping. By leveraging nonlinear optical computations with microwatts of average optical power, the researchers have made significant progress towards realizing the potential of optical neural networks.
The study demonstrated that a specific set of model weights can be selected from the weight bank in optics, negating the need for specialized computing hardware. This approach utilizes the natural phenomena of light propagation and eliminates the effort, cost, and complexity of manufacturing and operating dedicated devices for this purpose.
The implications of this research are far-reaching, particularly in addressing the challenges posed by the increasing 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.
The computational framework showcased in the experiment can be extended to efficiently program various high-dimensional, nonlinear phenomena for performing a wide range of machine learning tasks. This presents a transformative solution to the resource-intensive nature of current AI models.
The EPFL researchers’ work signifies a significant stride forward in the development of neural networks. By utilizing the unique properties of light and multimode fibers, they have unlocked the potential for more sustainable and efficient artificial intelligence systems. This breakthrough not only reduces the resource requirements but also contributes to the ongoing efforts to optimize computing hardware and machine learning algorithms for the advancement of AI.
*Note:
1. Source: Coherent Market Insights, Public sources, Desk research
2. We have leveraged AI tools to mine information and compile it
