Coherent optical neuron control based on reinforcement learning.

Area of Science:

• Optoelectronics
• Artificial Intelligence
• Computational Neuroscience

Background:

• Optical neural networks (ONNs) utilize optical neurons for complex computations.
• Coherent optical neurons are a key implementation, adept at natural and complex number calculations.
• Traditional control methods are inadequate due to the state variability and reliability demands of coherent optical neurons.

Purpose of the Study:

• To introduce a new control method for coherent optical neurons.
• To address the limitations of existing control strategies in ONNs.
• To demonstrate the effectiveness of the proposed control approach.

Main Methods:

• Development and experimental validation of deep reinforcement coherent optical neuron control (DRCON).
• Comparison of DRCON with standard stochastic gradient descent (SGD) for performance evaluation.

Main Results:

• DRCON demonstrated an average convergence rate 33% faster than SGD.
• The effective number of bits processed increased from under 2 to 5.5 bits with DRCON.
• Experimental results confirm the effectiveness of the proposed deep reinforcement learning control strategy.

Conclusions:

• DRCON provides a robust and efficient control mechanism for coherent optical neurons.
• The method significantly enhances computational performance metrics in optical neural networks.
• DRCON represents a crucial advancement towards the development of large-scale optical neural networks.