Efforts of optical physics researchers investigating the coherent Ising machine (CIM) are leading to breakthroughs, including uncovering new strategies for improving optimization computational abilities.

In his talk at Upgrade 2021, Dr. Hiroki Takesue, Senior Distinguished Researcher at NTT Basic Research Labs, discussed the updated status of the CIM being developed in Japan. Takesue highlighted three ways the CIM has evolved: implementing a magnetic field term, simulating 2D Ising models on the CIM, and spiking neural networks.

While previous versions of the CIM lacked a magnetic field term, Dr. Takesue’s team found implementing one drastically improved optimization mapping.

“With the strong magnetic field term, the station completely changes,” Takesue said. “The starting point is almost the same, but those speeds are forced to align into the same direction. What we obtain is that the magnetization is almost one and the spin-spin correlation is almost one.”

The same magnetization and spin-spin correlation led to a clear crossover of an anti-ferromagnetic state governed by the spin-spin interaction to the ferromagnetic state forced by the external field. Further, the resulting curve aligns closely with the hypothesized analytical solution.

In the case of simulating a 2D Ising model on the CIM, the goal was to improve CIM phase transitions. To achieve this, the researchers changed degenerate optical parametric oscillators (DOPOs) to increase the magnetization.

“Our purpose is to obtain this coupled experiment,” Takesue said. “We changed the coupling between DOPOs. Then, we assume the energy distribution obtained by the CIM with canonical distribution.” 

The result was phase transition-like behavior similar to the analytical solutions derived from theory.

Finally, Takesue discussed the application of simulating spiking neurons by using DOPOs.

Neurons are divided into two groups when it comes to reacting to thresholds. Class I neurons spike at a very low frequency. As the input strength increases, so does the spin frequency. When class II neurons reach their threshold, however, they tend to spike drastically and not change much going forward regardless of increases in pump strength.

Researchers paired DOPOs to artificially spiking neurons and found the system underwent opposite-sign optical coupling between two DOPOs, which led to a spiking signal by energy transfer between DOPOs.

“Automatically adaptive behavior is obtained in a spiking neuron,” Takesue said. “That is the current hypothesis, but of course this is one experiment using one particular instance. We need to do more investigation.”

All of these research findings bring Takesue and other researchers closer to developing the LASOLV model, short for laser solver, a computational system based on the coherent Ising machine.

“We expect that LASOLV will play an important role as one of the next generation of computers for the IOWN concept that is assumed by NTT,” Takesue said.