Conservation laws dictate memory in physical learning rules

A new research paper explores how conservation laws influence the memory and learning capabilities of physical learning rules like equilibrium propagation (EP) and coupled learning (CL). The study demonstrates that these rules conserve a property called "conductance mass," which stabilizes training and dictates the inductive bias. While EP and CL are shown to be trajectory equivalent for single outputs, adjoint coupled learning (AL) dissipates this mass. The research indicates that conservation laws play a crucial role in determining a physical learning machine's initialization memory, training speed, and generalization performance. AI

IMPACT Provides theoretical insights into the fundamental mechanisms of learning in physical systems, potentially influencing future hardware-based AI development.

RANK_REASON Research paper detailing theoretical findings on physical learning rules. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Adjoint Coupled Learning
• arXiv
• Conductance Mass
• cs.LG
• Equilibrium Propagation

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →