# Pavlov’s experiment-inspired optical neural networks based on dual-color fluorescence switching effect

## TL;DR

This paper introduces a new way to train optical neural networks using light, inspired by how biological systems learn, enabling faster and cheaper fabrication.

## Contribution

A novel 'top-down' in-situ training method for optical neural networks using dual-color fluorescence switching, inspired by Pavlovian associative learning.

## Key findings

- Sequential light exposure induces fluorescence switching and encodes memory in a dual-color resin.
- The framework successfully recognizes letters and can be extended to handwritten digit recognition.
- The method eliminates the need for weight computation in optical neural network fabrication.

## Abstract

Cutting-edge optical neural networks are often still trained by backpropagation, which is computationally intensive and originally for conventional artificial neural networks. Inspired by Pavlov’s experiment, we drew upon the principles of biological memory to establish an associative learning framework for training optical neural networks that mimics the mechanisms of associative learning and synaptic plasticity using dual-wavelength stimuli (i.e. ultraviolet and visible light) on a dual-color photoinitiator resin. Sequential light irradiation was shown to induce fluorescence switching and encode associative memory in the resin, which can serve as a physical substrate for an optical neural network. In optical experiments, the established framework was applied to pattern recognition of the letters ‘N,’ ‘V,’ and ‘Z.’ Simulations were conducted that extended its application to the recognition of handwritten digits. Compared to the current mainstream ‘bottom-up’ optical neural network fabrication approach that requires ‘weight calculation followed by hardware implementation’, this work presents a novel ‘top-down’ in-situ training methodology that eliminates the need for weight computation. The proposed method holds significant implications for large-scale, low-cost, and rapid fabrication of optical neural networks intended for edge computing applications. This study bridges biological learning principles with optical neural networks to provide a foundation for next-generation adaptive and scalable artificial intelligence systems.

Inspired by Pavlov’s experiment and associative learning theory, we realized in-situ hardware training of optical neural network with dual-color photoinitiator materials.