[10a-B21-1~11] FS.1 Focused Session "AI Electronics"

9:00 AM - 9:15 AM JST(12:00 AM - 12:15 AM UTC)

[10a-B21-1] Excitability learning of photonic spiking neural networks for combinatorial optimization

〇Takahiro Inagaki¹, Kensuke Inaba¹, Yasuhiro Yamada¹, Toshimori Honjo¹, Takuya Ikuta¹, Yuya Yonezu¹, Takushi Kazama², Koji Enbutsu², Takeshi Umeki², Kazuyuki Aihara³, Hiroki Takesue¹ (1.NTT BRL, 2.NTT Device Technology Laboratories, 3.The University of Tokyo)

Keywords:

Spiking neural network,Degenerate optical parametric oscillator,Neuromorphic computing

Spiking neural networks (SNNs) emulate the firing dynamics of biological neurons, where both synaptic learning and neuronal excitability play essential roles in computation. We previously demonstrated Class-I and Class-II firing modes using degenerate optical parametric oscillators (DOPOs) and applied a large-scale optical SNN with up to 10,240 nodes to combinatorial optimization problems. A unique feature of our optical neurons is their ability to autonomously adjust firing frequencies according to network conditions, preferentially activating poorly performing nodes. In this work, we propose a learning rule that updates the intrinsic excitability of each neuron based on observed firing frequencies during optimization. Experimental results show that the proposed method improves solution accuracy.