Tensor operations form the basis of almost all modern technologies, especially artificial intelligence, and go beyond the basic mathematics we know. In basic terms, tensors would be like atomic bricks in modern computing, especially AI. And tension operations are the “tools” (hammers, saws, screws) that we use to assemble, sculpt and transform these bricks into complex and useful structures, like models that recognize our voice, recommend movies or drive autonomous cars.
Today, all AI tasks, from image recognition to natural language processing, depend on tensor operations. However, the explosion of data has pushed conventional digital computing platforms, such as GPUs, to their limits in terms of speed, scalability, and power consumption. While humans and conventional computers must perform these operations step by step, light can perform them all at once.
To solve this problem, an international team of scientists, led by Yufeng Zhang from the Photonics Group of the Department of Electronics and Nanoengineering at Aalto University, has discovered a new method that performs complex computations with tensors using a single propagation of light. The result is tensor computing in a single pass, at the speed of light. The results have been published in Nature.
“Our method performs the same operations as current GPUs, but at the speed of light – says Zhang -. Instead of relying on electronic circuits, we use the physical properties of light to perform multiple calculations simultaneously.
To achieve this, Zhang’s team encoded digital data into the amplitude and phase of light waves, thereby converting the numbers into physical properties of the optical field. When these light fields interact and combine, they naturally perform mathematical operations such as multiplication. of matrices and tensors, which constitute the basis of deep learning algorithms. By introducing multiple wavelengths of light, the team extended this approach to handle higher-order tensor operations.
“Imagine you are a customs officer who has to inspect each package through several machines with different functions and then sort them into the correct containers – explains Zhang – Normally, you would process each package one by one. Our optical computing method integrates all packages and all machines– We create multiple optical links connecting each input to its correct output. With a single operation, a single pass of light, all inspections and classifications are carried out instantly and in parallel.”
Another key advantage of this method is its simplicity. Optical operations occur passively as light propagatesso no active control or electronic switching is needed during calculation.
“This approach can be implemented on almost any optical platform – adds Zhipei Sun, leader of the Photonics Group at Aalto University -. In the future, we plan to integrate this computational framework directly into photonic chips, which will allow processors based in light perform complex AI tasks with an energy consumptionYoto extremely low”.
Ultimately, the goal is to implement the method on existing hardware or platforms of large companies, say the authors, who conservatively estimate that the approach will be integrated into these platforms within 3 to 5 years.
“This will create a new generation of optical computing systems, significantly accelerating complex AI tasks in a multitude of fields”, concludes the study.