Researchers from the University of Oxford, in collaboration with colleagues from the Universities of Münster, Heidelberg, and Ghent, have made a surprising discovery: using less complex light sources can actually improve the performance of some optical applications, particularly light-driven AI technologies. This finding, published in the journal *Nature*, could lead to cheaper and more energy-efficient systems for the future of artificial intelligence.
The team's research centred around the concept of coherence in light sources. Coherence refers to the consistency of light waves in both time and space. Highly coherent sources, such as lasers, emit light within a narrow range of wavelengths, producing a single colour. This high level of coherence is crucial for many modern technologies, including optical communications, LiDAR, and medical imaging.
However, the study challenges this conventional wisdom by demonstrating that partially coherent light sources, which emit light across a broader spectrum, can outperform lasers in certain applications. Specifically, they found that using partially coherent light in photonic AI accelerators â devices that use photons instead of electrons for AI calculations â can significantly enhance parallel processing capabilities.
The researchers utilised a partially coherent light source generated from a narrow portion of the spectrum of incoherent light produced by an erbium-doped fibre amplifier, a device commonly used in optical communication to amplify light signals. This light was then split and distributed across different input channels within a parallel AI computational array. The results were remarkable: the parallelism of AI computation was boosted by a factor of N in a photonic accelerator with N input channels.
To illustrate the potential of this discovery, the team applied their system to identify Parkinson's disease patients through gait analysis, achieving a classification accuracy of over 92%. They further demonstrated how a simple system with only one partially coherent light source and nine input channels could perform high-speed AI tasks at a rate of 100 billion operations per second. Such speeds, equivalent to playing more than two hours of 4K video in a single second, would typically require multiple separate coherent lasers in a conventional photonic AI accelerator.
The implications of this discovery are significant. Dr. Bowei Dong, lead author of the study, highlights the scaling effect: "The benefit of using 'poorer' light sources has a scaling effect. You can run your AI models 100 times faster compared to a laser system, if the photonic accelerator scales to 100 input channels."
Professor Harish Bhaskaran, co-founder of Salience Labs and head of the research team, suggests that this breakthrough could extend beyond photonic computing: "While this work showcases the use of partially coherent light in emerging areas of photonic computing, we will in future also investigate whether this insight might apply to optical communications, particularly in the emerging optical interconnect technology space."
This exciting research opens up new possibilities for developing more efficient and affordable AI technologies, with potential applications ranging from healthcare to communications. The future of photonic computing appears brighter than ever, thanks to this unexpected discovery about the power of partially coherent light.
