We present a proof-of-concept demonstration of an all-optical inference architecture for classification of handwritten digits using structured Bessel-like beams. By employing a single-plane phase-modulation scheme controlled by only seven trainable parameters, our system steers the beam’s intensity into class-specific regions at the detection plane. These parameters are optimized via gradient descent, enabling direct physical training of the optical platform. In experiments on the MNIST dataset, the system achieves nearly $75\mathrm{\%}$ classification accuracy, underscoring the feasibility of harnessing structured light and minimal hardware complexity for optical machine-learning tasks. This work highlights how the inherent properties of Bessel-like and structured beams can be leveraged to perform classification, offering a pathway toward compact and efficient all-optical computing.

Close

PrevNext

View figure in article

• Analog computation
• Angular momentum of light
• Geometrical & wave optics
• Imaging & optical processing
• Interference & diffraction of light
• Non-diffracting waves
• Structured light
• Optical computing