NTT Research and Tohoku University Reveal Single-Photon CIM
NTT Research
Quantum leap: Pioneering research shows single-photon coherent ising machines outperform expectations
In the Quantum Science and Technology journal, NTT Research and Tohoku University's Graduate School of Information Science (GSIS) announced groundbreaking quantum computing and optimisation findings. According to their publication, “Single photon coherent Ising machines for constrained optimisation problems,” single-photon coherent Ising machines (CIMs) can outperform traditional CIMs, even those with higher photon counts, in solving difficult combinatorial optimisation problems. Quantum simulation shows that quantum entanglement can improve performance even in extremely low light levels, challenging signal-to-noise constraints.
Coherent Ising Machine
Cooperation Speeds Cyber CIM Development
The 2023 JRA formalises NTT Research and Tohoku University's relationship. This partnership aims to create a large-scale CIM simulation platform using high-performance computing . This large-scale CIM aims to enable free “cyber CIMs” that tackle hard NP, NP-complete, and HP-hard problems.
According to this agreement:
Professor Hiroaki Kobayashi of GSIS at Tohoku University is the JRA's principal investigator, and PHI Lab director Yoshihisa Yamamoto of NTT Research is his counterpart.
Tohoku University research optimises third-generation cyber CIM using HPC systems. This comprises optimising cache memory architecture data management, vectorising and parallelising kernels as accelerators, and scaling the cyber CIM to 100 million spins with sparse connections on a suitable platform.
This agreement will “unlock energy efficient and optimised machine learning accelerators” and bring them closer to a large-scale CIM simulator, said NTT Research PHI Lab Director Yoshihisa Yamamoto.
The PHI Lab's main purpose, which drives the JRA, is to leverage nonlinear quantum optical technology to create simple, helpful computer devices for common problems. Reimagining analog/digital hybrid computers influenced by quantum physics and neuroscience, such brain computers, is required.
CIMs are Ising model-mapped networks of degenerate optical parametric oscillators (DOPOs). Programed oscillators solve combinatorial optimisation problems. A mathematical depiction of magnetic systems with competitive spins is the Ising model. Together, researchers from both schools study a combinatorial clustering problem, a common unsupervised machine learning task.
Striking and Unexpected Single-Photon CIM Results
One photon per pulse is average in the recently proposed CIM. This figure is eight orders of magnitude lower than the known photon number (10⁸ photons per pulse) in typical CIMs. In such a weak light limit, quantum theory must be used to evaluate CIM performance, not heuristics.
It was once thought that a CIM that used one photon per pulse would have problems storing analogue amplitude data and monitoring internal pulse amplitudes due to its low signal-to-noise ratio. Under this assumption, its performance was expected to be significantly lower than standard CIMs.
The quantum model's numerical simulations yielded a surprising and unique conclusion. The research team showed that the single-photon CIM worked better than photon-rich versions, finding more correct answers for varied problem scenarios.
Explaining Quantum Enhancement
Due to a quantum mechanical phenomena, the single-photon CIM performed better in this partnership.
The CIM's measurement port needs an extraction beam splitter to produce a correlated internal pulse and an extracted pulse for measurement.
This correlation between extracted and internal pulses approaches a quantum regime even with background noise. The single-photon CIM strongly shows that these two pulses are quantumly entangled.
The single-photon CIM showed that background noise and optical loss can break intrinsically weak quantum entanglement, but it can also create substantial classical correlations between measured pulses. This conversion occurs via quantum measurement and feedback.
Quantum entanglement and its instantaneous conversion to classical correlation improve the single-photon CIM's performance. Traditional CIMs, which use several photons per pulse, lack this process.
Future prospects and practical applications
NTT Research and Tohoku University want to deepen their collaboration. Based on the single-photon CIM's strong theoretical confirmation, their immediate goals include implementing it. They will also accelerate Cyber CIM, a huge simulation environment. This joint effort aims to “pave the way for fast and energy-efficient solutions to real-world industrial problems” and maybe apply to energy-efficient machine learning.