Silicon integrated photonic-electronic neuron for noise-resilient deep learning.

Area of Science:

• Photonics
• Artificial Intelligence
• Integrated Circuits

Background:

• Photonic-electronic multiply accumulate neuron (PEMAN) architectures offer potential for high-speed computation.
• Neural networks (NNs) require efficient hardware for complex tasks like health monitoring.
• Noise sensitivity in photonic integrated circuits can limit performance.

Purpose of the Study:

• To experimentally demonstrate the photonic segment of a PEMAN architecture.
• To evaluate its performance in a noise-sensitive NN for heartbeat sound classification.
• To introduce and validate noise mitigation strategies for photonic neural networks.

Main Methods:

• Utilized a silicon photonic chip with electro-absorption modulators for matrix-vector multiplication.
• Implemented a three-layer NN with 1350 trainable parameters.
• Employed quantization- and noise-aware deep learning techniques, including a novel activation function slope stretching strategy.
• Validated performance at compute rates of 10, 20, and 30 Gbaud.

Main Results:

• Achieved F1-scores of 85.9% at 10 Gbaud and 81% at 20 Gbaud for heartbeat classification.
• Demonstrated enhanced noise-resilient properties through simulations, showing excellent agreement with experimental data.
• Successfully mitigated noise impairments using the novel training model and activation function strategy.

Conclusions:

• The demonstrated photonic segment of PEMAN is effective for high-speed, noise-resilient neural network computations.
• The developed noise mitigation strategies are crucial for practical applications of photonic computing in health monitoring.
• This work paves the way for advanced photonic integrated circuits in AI-driven applications.