Temporal Contrastive Learning through implicit non-equilibrium memory

Area of Science:

• Computational neuroscience
• Machine learning
• Statistical physics

Background:

• Backpropagation is key for neural networks, but local learning offers decentralized training benefits for energy-based models.
• Local learning methods often contrast desired (clamped) and spontaneous (free) behaviors, but require explicit memory.

Purpose of the Study:

• Introduce Temporal Contrastive Learning (TCL) for energy-based models.
• Enable contrastive learning without explicit memory using implicit, non-equilibrium memory.
• Explore the role of non-equilibrium dissipation and energy costs in TCL.

Main Methods:

• Developed TCL using integral feedback for implicit memory.
• Employed a sawtooth-like temporal protocol alternating free and clamped behaviors during training.
• Analyzed learning quality and Landauer-like energy costs.

Main Results:

• TCL successfully generates implicit memory through integral feedback and temporal protocols.
• Non-equilibrium dissipation was shown to enhance learning quality.
• A Landauer-like energy cost for contrastive learning was determined.

Conclusions:

• TCL offers a novel approach for decentralized learning in energy-based models.
• Implicit non-equilibrium memory broadens the applicability of contrastive learning.
• Understanding energy costs is crucial for physical implementations of learning systems.