RF-photonic deep learning processor with Shannon-limited data movement.

Area of Science:

• Photonics and Artificial Intelligence
• Advanced Communication Systems
• Semiconductor Technology

Background:

• Edholm's law predicts exponential growth in communication data rates, necessitating new computing paradigms beyond Moore's Law.
• Deep neural networks (DNNs) face increasing computational demands, challenging current hardware accelerators.
• Optical neural networks (ONNs) offer potential for high-speed AI but face scalability and system overhead issues.

Purpose of the Study:

• To introduce a novel artificial intelligence hardware accelerator for advanced communication systems.
• To demonstrate a fully analog deep learning approach for processing raw radio frequency (RF) signals.
• To address the limitations of current ONNs in terms of scalability and system overhead.

Main Methods:

• Development and experimental validation of a multiplicative analog frequency transform optical neural network (MAFT-ONN).
• Implementation of fully analog deep learning computations directly on RF signals.
• Testing MAFT-ONN for modulation classification and MNIST digit classification tasks.

Main Results:

• MAFT-ONN achieved 95% accuracy in modulation classification tasks with rapid convergence.
• Demonstrated scalability with nearly 4 million fully analog operations for MNIST digit classification.
• Achieved speeds hundreds of times faster than traditional RF receivers due to analog data movement.

Conclusions:

• MAFT-ONN presents a promising solution for AI hardware acceleration in future communication systems like 6G.
• The analog processing approach overcomes limitations of digital RF receivers and current ONN architectures.
• This technology enables efficient, high-speed deep learning on raw RF signals, paving the way for next-generation wireless technologies.