Deep physical neural networks trained with backpropagation.

Area of Science:

• Artificial Intelligence
• Materials Science
• Physics

Background:

• Deep-learning models are essential in science but face scalability limits due to high energy demands.
• Current deep-learning accelerators primarily focus on energy-efficient inference, not training unconventional hardware.
• A major impediment is the inability to apply backpropagation to train novel physical hardware in situ.

Purpose of the Study:

• Introduce a hybrid algorithm, physics-aware training, to enable backpropagation for training physical systems.
• Demonstrate the training of deep physical neural networks using controllable physical substrates.
• Showcase the universality of the approach across optics, mechanics, and electronics.

Main Methods:

• Developed a hybrid in situ-in silico algorithm named physics-aware training.
• Applied backpropagation to train deep physical neural networks composed of controllable physical systems.
• Utilized diverse physical substrates including optics, mechanics, and electronics.

Main Results:

• Successfully trained diverse physical neural networks using physics-aware training.
• Demonstrated experimental audio and image classification tasks with these physical neural networks.
• Showcased the combination of backpropagation's scalability with in situ algorithm's noise mitigation.

Conclusions:

• Physics-aware training enables the use of backpropagation for training physical neural networks.
• Physical neural networks offer potential for faster, more energy-efficient machine learning.
• This approach can imbue physical systems with automatically designed functionalities for various applications.