- 2018.2.26An English version of both the QNNcloud Service Terms of Use and the Privacy Policy has been published. In addition, the Japanease version of the QNNcloud Terms of Use has been revised.
- 2018.2.22An English version of the QNNcloud Service Terms of Use and the Privacy Policy will be published on Feb. 26, 2018. With this publication, the Japanease version of the QNNcloud Terms of Use will be revised and incorporated in the provisions as Article 22(Language).
- 2018.2.16The QNNcloud Terms of Use were scheduled to be revised on Feb. 16, 2018. However, the revision has been postponed. We will announce a new date for the revision later.
- 2018.2.9The QNNcloud Terms of Use will be revised on Feb. 16, 2018. With this revision, an English version of the QNNcloud Terms of Use and the Privacy Policy will be published and incorporated in the provisions as Article 22.
- 2017.11.27 QNNcloud is now opened.
- 2017.11.27 QNN simulator will be coming soon.
What is QNNcloud?
Quantum neural networks (QNNs) are novel computing systems that use optical parametric oscillators (OPOs) as quantum neurons and optical homodyne measurement-feedback circuits as quantum synapses. The QNN efficiently searches for a solution for various combinatorial optimization problems by exploiting collective symmetry breaking at the OPO threshold. Users can experience what it is like to conduct real experiments with the QNN and numerical simulations based on the quantum theory of OPO networks. In the future, the QNNcloud will provide a simulation tool specifically for the development of algorithms for real-world applications.
The QNNcloud is based on a network of 2000 OPOs with programmable all-to-all connections so that users can solve NP-hard Max-Cut problems up to the problem size of N = 2000 on complete graphs without the effort of imbedding a target graph in the hardware.
Targets of QNNcloud
Problems involving combinatorial and continuous optimization are ubiquitous in our modern life. Classic examples include lead compound optimization in the development of medicine, frequency band and/or transmission power optimization in wireless communications, sparse coding for compressed sensing, Boltzmann sampling in machine learning, portfolio optimization in Fin Tech, and many others. Most of those problems belong to the non-deterministic polynomial (NP), NP-complete, and NP-hard classes in complexity theory and require exponential resources as the problem size increases. The QNN exploits quantum parallel searching at below OPO threshold, collective symmetry breaking at OPO threshold, and exponential probability amplification of the solution state at above OPO threshold to address this limitation.
QNN’s Hardware
In a 1-km long fiber ring cavity, N = 2000 OPO pulses are simultaneously generated by exciting an intra-cavity and periodically poled LiNb03 (PPLN) waveguide device using a pump pulse train with a 1-GHz repetition frequency. A binary variable is represented by the bi-stable 0-phase and π-phase states of each OPO pulse. Each OPO pulse is prepared in a 0-phase and π-phase superposition at below threshold but with either one of the two at well above threshold. Any pair of OPO pulses can be coupled by sequentially measuring the OPO pulse amplitude, computing an appropriate feedback pulse amplitude with a field programmable gate array (FPGA), and injecting the feedback pulse into the target OPO pulse. All-to-all connections for N = 2000 OPO pulses are implemented at every round trip (lasting 5 μsec). If the external pump rate increases to above OPO threshold, the exact solution or approximate solution is obtained as an oscillation phase (0 or π-phase) configuration after 10 to 1000 round trips.external pump rate increases to above OPO threshold, the exact solution or approximate solution is obtained as an oscillation phase (0 or π-phase) configuration after 10 to 1000 round trips.
QNN Simulator
The dynamics of the QNN can be theoretically predicted by the quantum master equation taking into account the wave packet reduction induced by measurements. One unique property of an OPO network is that the operation continually crosses from quantum limits to classical limits as computation progresses. The QNN simulator visually presents the differences between these two regions as well as the physics of this migration process from the quantum to the classical.
Future plans
We will release algorithms for various real-world applications, numerical simulation tools for the development of new algorithms, and a novel QNN with a recurrent neural network architecture in the coming years.