The international research team uses photonic networks to discover the models.
In the digital age, data traffic is growing at an exponential rate. For applications such as artificial intelligence, such as pattern and voice recognition, or for self-driving vehicles, the demand for computing power exceeds the capacity of conventional computer processors. Working with an international team of researchers University of Münster new approaches and process architectures are being developed that can deal with these tasks very effectively. They have now shown that so-called photonic processors, which process data with light, can process information much faster and in parallel – something they are not capable of making electronic chips. The results have been published in the journal Nature.
Background and methodology
Light-based processors that accelerate tasks in the machine learning field allow complex mathematical tasks to be processed at an enormous speed (10¹⁵² -10¹⁵ operation per second). Conventional chips like graphics cards or specialized hardware like Google’s TPU (Tensor Processing Unit) are based on electronic data transfer and are much slower. Wolfram Pernice Prof. The Institute of Physics and the research team at the Center for Nanoscience at the University of Münster implemented a hardware accelerator called matrix multiplication, which represents the main processing load in the calculation of neural networks. Neural networks are a set of algorithms that simulate the human brain. This is helpful, for example, in classifying objects in pictures and recognizing speech.
The researchers combined photonic structures with phase change materials (PCMs) as energy-efficient storage elements. PCMs are often used with DVD or BluRay discs to store optical data. The new processor allows the elements of the array to be stored and conserved without the need for a power supply. To perform matrix multiplications in parallel in multiple data sets, physicists in Münster used a chip frequency comb as a light source. The frequency comb offers a number of optical wavelengths that are processed independently of each other on the same photonic chip. As a result, this allows data to be processed in parallel by calculating all wavelengths at once – also called wavelength multiplexing. “Our study is the first to artificially apply frequency combs in the field of neural networks,” says Wolfram Pernic.
In the experiment, physicists used a so-called neural network to detect handwritten numbers. These networks are a concept in the field of machine learning inspired by biological processes. They are mainly used in image or audio data processing, as they currently achieve the highest classification accuracy. “The convulsive operation between input data and one or more filters — such as highlighting the edges of a photo — can very well be transferred to the architecture of our array,” explains lead author Johannes Feldmann. “Exploiting the light to transfer the signal allows the processor to perform parallel data processing by wavelength multiplexing. This results in a higher computational density and many matrix multiplications are performed over a single period of time. Compared to traditional electronics that typically work in the low GHz range, optical modulation speeds can be achieved at speeds ranging from 50 to 100 GHz. “This means that the process allows for data speeds and computational densities, which are operations on every surface of the processor that have never been achieved.
The results have many applications. In the field of artificial intelligence, for example, more data can be processed simultaneously while saving energy. The use of larger neural networks allows for more accurate and unattainable predictions and more accurate data analysis. For example, photonic processors help to evaluate large amounts of data in medical diagnostics, such as high-resolution 3D data generated by special imaging methods. More applications are in the areas of self-driving vehicles, such as rapid and rapid evaluation of sensor data and storage infrastructure, such as computing power or IT infrastructures that provide application software.
For more information on this study, read the AI driven by Light-Based Parallel Convolutions Processors.
Reference: J. Feldmann, N. Youngblood, M. Karpov, H. Gehring, X. Li, M. Stappers, M. Le Gallo, X. Fu, A. Lukashchuk, “Parallel processing through the nucleus of the integrated photonic voltage”. , AS Raja, J. Liu, CD Wright, A. Sebastian, TJ Kippenberg, WHP Pernice and H. Bhaskaran, January 6, 2021, Nature.
DOI: 10.1038 / s41586-020-03070-1
Research partners: Researchers from the University of Münster, Oxford and Exeter University in England, Pittsburgh University in the USA, École Polytechnique Fédérale (EPFL) in Lausanne (Switzerland) and IBM Research Laboratory in Zurich. they also took part in that work.
Funding: The research received funding from the EU’s “FunComp” project and the European Research Council (ERC grant “PINQS”).