Neuromorphic photonics with electro-absorption modulators.

Area of Science:

• Photonics
• Optical Computing
• Artificial Intelligence

Background:

• Photonic neural networks leverage light's properties for high-capacity computation.
• Integrated photonic components enable monolithic multi-layer networks, reducing energy and latency.
• Absorption modulators offer a more compact alternative to interferometer-based modulators for neural network density.

Purpose of the Study:

• To develop and model absorption modulators for electro-optic fully connected neural networks.
• To compare the performance of different absorption modulators as activation functions.
• To assess the impact of noise on network performance.

Main Methods:

• Developed a model for absorption modulators in electro-optic neural networks, incorporating noise.
• Evaluated five types of absorption modulators for their activation function properties.
• Tested a 2-hidden-layer feed-forward photonic neural network on MNIST classification.

Main Results:

• The quantum well absorption modulator demonstrated the best performance among the tested types.
• Achieved 96% prediction accuracy on MNIST classification.
• Reported an energy efficiency of 1.7×10^-12 J/MAC (excluding laser power).

Conclusions:

• Quantum well absorption modulators are highly effective for electro-optic neurons in photonic neural networks.
• Absorption modulators provide a viable and efficient approach for on-chip optical machine learning.
• This work advances the development of dense and energy-efficient photonic neural network hardware.