Minimum perturbation theory of deep perceptual learning

Area of Science:

• Neuroscience
• Computational Neuroscience
• Statistical Physics

Background:

• Perceptual learning (PL) demonstrates lasting improvements in task performance after training, linked to neural plasticity in sensory areas.
• Understanding the precise synaptic changes and dynamics of PL is crucial for connecting learning to neural plasticity.
• The distributed nature of learning-related changes across sensory hierarchies presents a theoretical challenge.

Purpose of the Study:

• To model perceptual learning (PL) in a deep nonlinear neural network representing the sensory hierarchy.
• To develop a statistical mechanics-based mean-field theory for PL of fine discrimination.
• To investigate the necessity and sufficiency of synaptic weight modifications for PL and explore normative principles governing these changes.

Main Methods:

• Modeled the sensory hierarchy as a deep nonlinear neural network.
• Applied statistical physics tools to derive a mean-field theory in the thermodynamic limit (large neurons, large examples).
• Characterized the solution space and introduced a minimum perturbation (MP) principle to constrain learning dynamics.

Main Results:

• The network's input-output function maps to a deep linear network in the thermodynamic limit.
• Synaptic weight modifications in the first layer are both sufficient and necessary for PL.
• The MP principle explains how plasticity in all layers, not just the first, can arise and reduce overall network perturbation, aligning with experimental findings.

Conclusions:

• A statistical mechanics framework provides mechanistic and normative insights into PL.
• PL dynamics are constrained by a minimum perturbation principle, favoring minimal synaptic changes.
• This model reconciles empirical observations of PL with underlying neural plasticity mechanisms.