Frequency Propagation: Multimechanism Learning in Nonlinear Physical Networks.

Area of Science:

• Physics
• Machine Learning
• Network Science

Background:

• Traditional machine learning algorithms often require centralized processing and global information.
• Physical networks, such as electrical or flow networks, offer unique opportunities for distributed computation and learning.
• Developing efficient, local learning rules for physical systems is crucial for their practical application.

Purpose of the Study:

• To introduce frequency propagation, a novel learning algorithm for nonlinear physical networks.
• To demonstrate that frequency propagation enables gradient descent on a loss function.
• To establish frequency propagation as a type of multimechanism learning strategy applicable to various physical networks.

Main Methods:

• Applied an activation current at one frequency and an error current at another frequency to a resistive electrical circuit with variable resistors.
• Analyzed the voltage response as a superposition of activation and error signals in the frequency domain.
• Updated circuit conductances locally, proportionally to the product of coefficients derived from the activation and error signals.

Main Results:

• Frequency propagation successfully updates network parameters using local information.
• The learning rule was proven to perform gradient descent on a defined loss function.
• Demonstrated the algorithm's applicability to nonlinear physical networks, including resistive, elastic, and flow networks.

Conclusions:

• Frequency propagation is an effective and local learning algorithm for physical networks.
• The algorithm exemplifies a multimechanism learning strategy, utilizing distinct physical quantities as signals.
• This approach offers a unified framework for understanding various physical network learning mechanisms, including prior work in flow networks.