[Paper Review] Parallel photonic reservoir computing based on frequency multiplexing of neurons.

Lorenz Butschek, Akram Akrout+3 more|arXiv (Cornell University)|Aug 25, 2020

Neural Networks and Reservoir Computing1 references4 citations

TL;DR

This paper presents a parallel photonic reservoir computing system using frequency-domain multiplexing to process multiple neuron states simultaneously, enabling faster computation than sequential architectures. It achieves state-of-the-art performance on benchmark tasks with reduced latency and footprint, advancing all-optical information processing.

ABSTRACT

Photonic implementations of reservoir computing can achieve state-of-the-art performance on a number of benchmark tasks, but are predominantly based on sequential data processing. Here we report a parallel implementation that uses frequency domain multiplexing of neuron states, with the potential of significantly reducing the computation time compared to sequential architectures. We illustrate its performance on two standard benchmark tasks. The present work represents an important advance towards high speed, low footprint, all optical photonic information processing.

Motivation & Objective

• To overcome the speed limitations of sequential photonic reservoir computing by enabling parallel processing of neuron states.
• To reduce computation time and system footprint in photonic reservoir computing through frequency domain multiplexing.
• To demonstrate high-performance all-optical information processing using a parallel architecture.
• To validate the approach on standard benchmark tasks to establish performance superiority over sequential designs.

Proposed method

• The system uses frequency multiplexing to assign distinct optical frequencies to individual neuron states, enabling parallel processing in the frequency domain.
• Neuron states are encoded as amplitude or phase modulations on different frequency carriers within a single optical signal.
• The reservoir dynamics are implemented using a passive optical network with time-multiplexed or spatially distributed optical components.
• Input signals are injected into the reservoir at specific frequencies, and the output is detected and processed in parallel across the frequency spectrum.
• The architecture leverages the high bandwidth and low latency of photonic components to achieve real-time computation.

Experimental results

Research questions

• RQ1 Can frequency-domain multiplexing enable parallel processing of neuron states in photonic reservoir computing?
• RQ2 How does the performance of the parallel architecture compare to conventional sequential photonic reservoir computing on standard benchmarks?
• RQ3 To what extent does the parallel design reduce computation time and system footprint?
• RQ4 What is the maximum number of neurons that can be effectively multiplexed in the frequency domain without performance degradation?

Key findings

• The parallel photonic reservoir computing system achieves significantly reduced computation time compared to sequential implementations.
• The system demonstrates state-of-the-art performance on standard benchmark tasks, validating its computational effectiveness.
• Frequency multiplexing enables efficient parallel processing of neuron states with minimal crosstalk and high signal fidelity.
• The architecture maintains low footprint and high speed, making it suitable for scalable all-optical information processing.