Unsupervised pretraining in biological neural networks.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Perceptual learning in the sensory cortex involves neural plasticity.
• The role of supervised versus unsupervised learning in this plasticity is not well understood.
• Representation learning in neural networks can be supervised or unsupervised.

Purpose of the Study:

• To determine whether neural plasticity in the visual cortex during learning is driven by supervised or unsupervised mechanisms.
• To investigate the role of different visual areas (V1, HVAs) in this process.
• To explore the potential impact of unsupervised learning on subsequent task learning.

Main Methods:

• Simultaneous recording of up to 90,000 neurons from the primary visual cortex (V1) and higher visual areas (HVAs) in mice.
• Comparing neural activity during task learning with unrewarded exposure to the same stimuli.
• Analyzing neural plasticity patterns in relation to visual and spatial learning rules.
• Behavioral experiments to validate predictions about unsupervised learning's effect on task acquisition.

Main Results:

• Neural changes during task learning were largely replicated by unrewarded stimulus exposure, indicating unsupervised learning.
• Neural plasticity was most pronounced in medial higher visual areas (HVAs) and followed visual, not spatial, learning rules.
• A reward-prediction signal was observed in anterior HVAs exclusively in task-performing mice, suggesting a role in supervised learning.
• Unsupervised learning was found to accelerate subsequent task learning.

Conclusions:

• Unsupervised learning is a dominant mechanism underlying neural plasticity in the mouse visual cortex during perceptual learning.
• Medial HVAs are key areas for visual, unsupervised learning, while anterior HVAs may support supervised learning via reward prediction.
• Unsupervised learning can prime the brain, potentially accelerating future supervised learning tasks.