Self-supervised predictive learning accounts for cortical layer-specificity.

Area of Science:

• Computational neuroscience
• Systems neuroscience
• Machine learning applications in neuroscience

Background:

• The neocortex's internal representation of the world is crucial but poorly understood.
• Existing models lack clarity on the circuitry and computational principles involved.

Purpose of the Study:

• To propose a computational theory for neocortical function inspired by self-supervised learning.
• To explain how layer 2/3 (L2/3) integrates sensory input and context for prediction.
• To elucidate the role of prediction errors in learning and behavior.

Main Methods:

• Developed a computational model where L2/3 predicts sensory input using past data and top-down context.
• Implemented self-supervised learning by comparing L2/3 predictions with L5 sensory representations.
• Validated the model using context-dependent temporal tasks and a sensorimotor task in silico.

Main Results:

• The model accurately predicts sensory information in complex tasks, showing robustness to noise and occlusion.
• Generated layer-specific sparsity patterns consistent with experimental findings.
• Model's prediction errors mirrored neural mismatch responses in awake mice during a sensorimotor task.

Conclusions:

• The multi-layered neocortex facilitates self-supervised predictive learning, enabling the brain to anticipate sensory stimuli.
• The proposed model offers testable predictions for the computational roles of cortical features.
• Findings suggest predictive coding as a fundamental principle of neocortical computation.