Resonant Machine Learning Based on Complex Growth Transform Dynamical Systems.

Area of Science:

• Electrical Engineering
• Machine Learning
• Optimization Theory

Background:

• Traditional energy-based learning models are inherently dissipative, associating a single energy metric with variable configurations.
• These models link the lowest energy state to optimal configurations, leading to energy loss during learning.

Purpose of the Study:

• To propose an energy-efficient learning framework by drawing parallels between machine learning and electrical networks.
• To develop a novel approach that minimizes energy dissipation during the machine learning process.

Main Methods:

• Exploiting structural and functional similarities between machine learning and electrical networks satisfying Tellegen's theorem.
• Associating active and reactive energy components with the network, ensuring zero reactive power in steady state.
• Utilizing a dynamical system with complex-domain, continuous-time growth transforms and an annealing procedure for optimization.

Main Results:

• The proposed framework dissipates active power only during learning, while reactive power remains zero.
• Learned parameters are self-sustained by electrical resonance, determined by network inductances and capacitances.
• Resonant support vector machines (SVMs) were designed, demonstrating reduced active power dissipation compared to non-resonant counterparts.

Conclusions:

• The novel framework enables energy-efficient learning by leveraging electrical resonance principles.
• This approach offers a new perspective on regularization and optimization in machine learning.
• The proposed method shows significant potential for designing power-efficient machine learning models.