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

9:30 AM - 9:45 AM JST(12:30 AM - 12:45 AM UTC)Presentation by Applicant for JSAP Young Scientists Presentation Award

[10a-B21-3] Brain-inspired computing based on photonic hyperdimensional vector generation

〇Takuya Iwata¹, Tomoaki Niiyama¹, Satoshi Sunada¹ (1.Kanazawa Univ.)

Keywords:

Optical Computing,Hyperdimensional Computing

Optical neural networks suffer from low noise tolerance and high optoelectronic-conversion power consumption. To address these issues, we propose a photonic implementation of hyperdimensional computing (HDC), a brain-inspired model. The phase of the input is modulated in a silicon photonic circuit and the light is spread through a scattering medium, generating optical high-dimensional vectors (HVs) in parallel as speckle patterns exceeding 100,000 dimensions. Learning is performed by bundling optical HVs into class HVs, and inference is realized by an all-photonic similarity calculation using a digital micromirror device (DMD): the target HV is projected onto the displayed class HVs, and the reflected intensity directly gives the similarity to classify the input. Since both encoding and inference rely only on light propagation, robust and energy-efficient computation is achieved. For MNIST (ten classes), the system attained about 91% accuracy with 160,000-dimensional optical vectors.