by Rohit Khorana (’24) | March 1, 2021
Sixty years ago, most people would have found the concept of self-driving cars to be something from a sci-fi story. Today, however, it is no longer a fantasy, with self-driving cars made possible through artificial intelligence (AI). AI is increasingly incorporated into the technology we use today, with important applications in areas such as healthcare and image recognition analysis. However, the increasing demand for artificial intelligence poses significant challenges for the computing hardware behind AI. Recently, an Australian team of researchers led by Xingyuan Xu and a German team led by Johannes Feldmann reported that photonic processors accelerate AI through the properties of light, a discovery that could completely revolutionize optical computing and AI.
Currently, conventional computing approaches in AI lag behind the speeds needed to support their use in everyday life. One common structure in AI is an artificial neural network (ANN). ANNs use many layers of interconnected artificial neurons to conduct complex mathematical operations, such as matrix-vector multiplication. To accelerate processing in these networks, scientists alter specific electronic computing systems that support the ANNs. Electrons are the primary carrier of information in these computing systems, but now, photons are an alternative and potentially an improvement.
Photons are a better option because of the properties related to their different wavelengths. Photons of different wavelengths can be multiplexed (transmitted in parallel) and modulated simultaneously without interference, minimizing the delay in terms of information signaling. However, there are some challenges associated with the use of photons. Currently, there are very few materials that support the speed of these new artificial neurons, and photonic devices have not yet been integrated into regular computing hardware.
Despite these challenges, advancements in optical frequency combs, light sources with sets of uniform spectra, have made integrated photonic processing possible. Xu’s team used optical frequency combs to accommodate different wavelengths of light simultaneously, enabling the researchers to achieve parallel computing of different types of operations. The speed is remarkable, clocking in at ten trillion operations per second, limited only by the amount of data fed into the system.
The team of researchers under Feldman independently came up with a different system also reliant on optical frequency combs. They integrated these combs into a computing architecture based on a phase-change material, a material that shifts from a crystalline to an amorphous phase. By combining the combs with the properties of the phase-change material, light-based operations are made even faster. In theory, this framework can accelerate operations to the speed of light with minimal energy consumption, which would be ideal for heavy-duty data processing such as that required for cloud computing.
Despite some issues with using photons to power artificial intelligence, the possible rewards are worth the challenges. Researchers in many different fields, including materials science, photonics, and electronics, are collaborating to build practical machinery to support photonic AI. The development of hardware that can successfully integrate photonic processors and electronic circuits could revolutionize AI. Shockingly, AI that operates almost as fast as the human brain is no longer fiction but reality.