Predictive learning rules generate a cortical-like replay of probabilistic sensory experiences

Area of Science:

• Computational Neuroscience
• Cognitive Neuroscience
• Neural Networks

Background:

• The brain constructs internal models of the environment's probabilistic structure.
• Spontaneous brain activity may reflect these internal models by replaying past sensory experiences.
• The neural mechanisms for encoding these models into spontaneous activity are not well understood.

Purpose of the Study:

• To investigate how recurrent neural networks can learn the spontaneous replay of probabilistic sensory experiences.
• To explore the role of soma-dendrite interactions in predictive coding within neural networks.
• To propose a novel mechanism for encoding internal models into spontaneous brain activity.

Main Methods:

• Developed a recurrent network of spiking neurons implementing a predictive learning principle.
• The learning rules minimized probability mismatches between evoked and internally driven neural activity.
• Analyzed the network's ability to generate stimulus-specific cell assemblies and replicate behavioral data.

Main Results:

• The model learned to replay probabilistic sensory experiences through internally driven neural dynamics.
• Stimulus-specific cell assemblies internally encoded activation probabilities via recurrent connections.
• The model's spontaneous activity successfully replicated behavioral biases observed in monkey perceptual decision-making tasks.

Conclusions:

• Recurrent networks with predictive learning can learn spontaneous replay of sensory experiences.
• Soma-dendrite interactions within neural networks are crucial for encoding internal probabilistic models.
• These findings highlight the importance of intracellular processes and network dynamics in cognitive learning.