Learning predictive signals within a local recurrent circuit.

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Machine Learning

Background:

• The predictive coding hypothesis posits that the brain compares top-down predictions with bottom-up sensory information, using prediction errors to signal discrepancies.
• Recent experimental evidence suggests prediction error signals may arise intrinsically within local neural circuits, independent of top-down input.

Purpose of the Study:

• To investigate whether local neural circuits alone can generate predictive signals.
• To model the emergence of prediction errors using local plasticity rules in a recurrent spiking network.

Main Methods:

• Training a recurrent spiking neural network model.
• Implementing local plasticity rules within the network.
• Analyzing the network's ability to replicate experimentally observed prediction error features.

Main Results:

• The recurrent spiking network model successfully replicated key features of experimentally observed prediction errors.
• Biphasic neural activity patterns characteristic of prediction errors were reproduced.
• Context dependency in prediction error signals was also replicated by the model.

Conclusions:

• Local circuits, through synaptic plasticity, can autonomously generate predictive signals and prediction errors.
• The findings support the idea that internal models of sensory input can be acquired and updated via local learning mechanisms.
• This work provides a computational framework for understanding the neural basis of prediction error generation.