Speaker
Description
Artificial intelligence (AI) has become widely adopted across many domains, but
conventional electronic implementations still suffer from high computational cost and
significant energy consumption during inference. Building on previously reported
optical neural networks for MNIST handwritten digit (0–9) classification (X. Lin, Y.
Rivenson et al., 2018), we simplify a diffractive architecture by using a visible laser
wavelength and a spatial light modulator (SLM) to assess its performance and
feasibility under a new optical configuration. The model is trained on a digital
computer, and the learned phase information is then encoded onto the SLM. By
moving to a shorter laser wavelength, we increase the achievable neuron density per
unit area, effectively enlarging the capacity of the optical neural network without
expanding its footprint. By adjusting the freespace propagation distances, we are able
to mimic the connections between neurons in successive layers. In future work, we
plan to extend this visiblewavelength optical neural network from classification tasks
to generative AI, targeting optical image generation on the same platform. This
progression would bridge optical classification and optical generative models, and
help quantify how spectral choices impact the quality, diversity, and energy efficiency
of optical platforms for AI generation.
