2023年第70回応用物理学会春季学術講演会

講演情報

一般セッション(口頭講演)

FS フォーカストセッション「AIエレクトロニクス」 » FS.1 フォーカストセッション「AIエレクトロニクス」

[15p-B414-1~11] FS.1 フォーカストセッション「AIエレクトロニクス」

2023年3月15日(水) 13:00 〜 16:00 B414 (2号館)

河口 研一(富士通)、松井 千尋(東大)

14:00 〜 14:15

[15p-B414-5] Increasing the Noise Margin of Ising Machine by Short-term Memory in Neuron-inspired Unit

Zhiqiang Liao1、Kaijie Ma1、Hiroyasu Yamahara1、Munetoshi Seki1、Hitoshi Tabata1 (1.Univ. of Tokyo)

キーワード:Ising machine, Combinatorial optimization, Stochastic resonance

The gain-dissipative Ising machine (GIM) is a high-speed dedicate solver for hard combinatorial optimization problems (COPs), which maps the COP to the Ising model and seeks the global optimum using the spontaneous descent of its Ising Hamiltonian. Under the interaction of bistable nonlinearity with linear gain dynamics, the output amplitudes of GIM bifurcate to model the two states of electron spin. However, recent reports revealed that traditional bistable GIMs require larger driving energies to suppress random thermal switching of spin state influenced by noise. This means that the traditional bistable GIM has a small noise margin, which limits further power-consumption reduction. Inspired by the property that neurons can use noise to enhance their own responses, we propose a nonlinear unit with short-term memory (SMNU) and study the performance of its hardware structure in simulation. It consists of some simple electronic components, including an integrator, adders, multipliers, and amplifiers. FPGA, which realizes spin interaction, forms a feedback loop with SMNU as a GIM.
In the present work, we show the similarity of FitzHugh-Nagumo neuron and the SMNU, which explains the short-term memory ability of SMNU. Then, by the domain clustering dynamics test, it is proved that the short-term memory ability of the SMNU enables the GIM to significantly suppress noise induced temporary domains compared with traditional GIMs. This implies that SMNU can achieve good performance even under a large noise environment. Further testing on some MAXCUT problems verifies this conclusion. Besides, it is revealed that the performance changing trend of the SMNU-based GIM with increasing noise is similar to the stochastic resonance effect in FitzHugh-Nagumo neurons.
In conclusion, we prove that SMNU-based GIM has a much larger noise margin compared with traditional ones, which gives it the potential to be a low-power-consumption candidate.