Human emotion recognition with a microcomb-enabled integrated optical neural network.

Area of Science:

• Optoelectronics
• Artificial Intelligence
• Neuromorphic Computing

Background:

• Deep learning models require significant computational power, driving demand for faster, low-power hardware.
• Current AI systems struggle with the energy and speed demands of complex tasks like emotion recognition.

Purpose of the Study:

• To propose and validate a microcomb-enabled integrated optical neural network (MIONN) for high-speed, low-power human emotion recognition.
• To demonstrate the potential of photonic-electronic computing for advanced AI applications.

Main Methods:

• Fabrication of proof-of-concept microcomb-enabled integrated optical neural network chips.
• Development of a photonic-electronic AI computing engine with automatic feedback control for stability and precision.
• Encoding large-scale tensor data using microcomb-generated frequency channels for parallel computation.

Main Results:

• The MIONN prototype achieved a potential throughput of 51.2 TOPS (tera-operations per second).
• The system demonstrated stable operation with 8-bit weighting precision.
• Successfully recognized six basic human emotions with 78.5% accuracy on a blind test set.

Conclusions:

• The proposed MIONN offers a high-speed and energy-efficient hardware solution for deep learning models.
• This technology enables AI with enhanced emotional interaction capabilities.
• Integrated optical neural networks represent a promising direction for next-generation AI computing.