Conservation Laws Determine Physical Learning Memory
A recent study published on arXiv, paper arXiv:2608.00097, explores physical learning rules including Equilibrium Physics (EP), Conservation Laws (CL), and Active Learning (AL). The research reveals that EP and CL maintain conductance mass in the small-nudge limit and are trajectory equivalent for single-output tasks. Conversely, AL dissipates conductance mass at double the rate of its loss. While conserved mass has no effect in linear circuits, the presence of nonlinear elements like diodes disrupts this protection. The findings advance understanding of inductive bias and have implications for physical neural networks and learning systems.
Key facts
- Paper arXiv:2608.00097 examines physical learning rules EP, CL, and AL.
- EP and CL conserve conductance mass in the small-nudge limit.
- EP and CL are trajectory equivalent for single-output tasks.
- AL dissipates conductance mass at twice its loss.
- In linear circuits, conserved mass has no functional consequence.
- Nonlinear elements like diodes break the protection.
- The paper proves theoretical results about inductive bias.
- The study is relevant to physical neural networks and learning systems.
Entities
Institutions
- arXiv