Modern models of artificial intelligence (AI) using billions of trained parameters are faced with the problem of huge costs for training and deployment. Models require a significant amount of memory and computing capacities available only in huge data centers that consume electricity, comparable to the needs of entire cities.
The research community is actively looking for ways to optimize both computing equipment and machine learning algorithms to maintain the development of AI. One of the promising areas is the optical implementation of neural networks architectures, which can significantly reduce energy consumption.
In the new study, published in the journal Advanced Photonics, researchers have explored the use of light spread in multimode fibers (MMF) to achieve the same level of effectiveness in image classification tasks as completely digital systems, but using over 100 times fewer parameters. This approach reduces memory requirements and decreases the need for energy-intensive computing processes, while maintaining high accuracy in various machine learning tasks.
A scheme of the experiment on the programming of optical distribution for a computing task was utilized. A Spatial Light Modulator (SLM) was used to modulate laser pulses with a data sample imposed on a fixed programming template. The bundle was connected to a multi-fold fiber. The resulting pattern after distribution was recorded by a camera. The classification accuracy of the output data was then calculated to assess the task’s accuracy, which was transmitted back to the surrogate optimization algorithm. The algorithm improves performance by exploring various programming parameters and refining potential solutions.
The basis of this work is precise control of ultra-short impulses in multimode fibers using the technique of forming a wavefront. This method enables non-linear optical calculations using micro-watts of medium optical power, which is a significant advancement in realizing the potential of optical neural networks.
Scientists have discovered that by using a small group of parameters, a specific set of models can be chosen from the ‘Bank of Libra’. In wider optical systems, the Bank of Libra is used to store and switch between different coefficients that determine the processing of signals. This allows for the optimization of systems for various applications, ranging from image processing to photon calculations, combining the principles of machine learning with optical technologies for efficient data processing.