Photonic deep residual time-delay reservoir computing.

Area of Science:

• Photonics
• Artificial Intelligence
• Machine Learning

Background:

• Time-delay reservoir computing (TDRC) offers a simplified approach to recurrent neural networks.
• Deep architectures can enhance TDRC capabilities but face challenges with layer redundancy.

Purpose of the Study:

• To propose and demonstrate a novel photonic deep residual TDRC (DR-TDRC) with augmented capabilities.
• To improve memory capacity and nonlinear channel equalization performance.
• To enable scalable deep TDRC architectures with enhanced stability.

Main Methods:

• Development of a photonic deep residual TDRC (DR-TDRC) architecture incorporating additional time delays in residual connections.
• Implementation of a specialized clipping algorithm to mitigate performance degradation in deep structures.
• Experimental demonstration using a 4-layer DR-TDRC with 960 neurons based on injection-locked distributed feedback lasers.

Main Results:

• DR-TDRC demonstrates superior performance over traditional deep structures in benchmark tasks.
• Achieved significant improvements in memory capability and nonlinear channel equalization (nearly an order of magnitude).
• Successfully extended deep TDRC to dozens of layers using the clipping algorithm, enhancing overall performance.

Conclusions:

• The proposed DR-TDRC offers a feasible and scalable approach for advanced deep photonic computing.
• This work addresses the limitations of deep TDRC, paving the way for more powerful AI hardware.
• The findings support the expansion of deep photonic computing to meet increasing artificial intelligence demands.