Perceptual learning as improved probabilistic inference in early sensory areas

Area of Science:

• Neuroscience
• Computational Neuroscience
• Perceptual Learning

Background:

• Perceptual learning significantly enhances performance on specific tasks.
• Existing models attribute this learning to either early visual cortex tuning or late decision-stage inference.
• These models fail to fully explain psychophysical and neurophysiological findings.

Purpose of the Study:

• To develop a unified model explaining both behavioral and neurophysiological aspects of perceptual learning.
• To investigate the role of feedforward connectivity in perceptual learning within recurrent neural networks.

Main Methods:

• Utilized a recurrent network of spiking neurons.
• Modified feedforward connectivity to enhance probabilistic inference in early visual areas.
• Simulated perceptual learning and analyzed network responses.

Main Results:

• The model successfully replicated both behavioral improvements and neurophysiological changes associated with perceptual learning.
• Observed modest alterations in tuning curves, consistent with experimental data.
• Demonstrated a significant reduction in pairwise noise correlations.

Conclusions:

• Altering feedforward connectivity in early visual areas can drive perceptual learning.
• This approach reconciles conflicting findings from previous early and late visual processing theories.
• The model provides a unified framework for understanding perceptual learning mechanisms.