Monkey Prefrontal Cortex Learns to Minimize Sequence Prediction Error.

Area of Science:

• Computational Neuroscience
• Cognitive Neuroscience

Background:

• The prefrontal cortex (PFC) is crucial for complex cognitive functions, including prediction and decision-making.
• Understanding the computational mechanisms underlying PFC function remains a significant challenge in neuroscience.

Approach:

• We introduce the Sequence Prediction Error Learning (SPEL) model, a novel recurrent neural network (RNN).
• SPEL minimizes sequence prediction errors by adjusting synaptic weights, drawing inspiration from deep learning principles used in large language models.
• Unlike standard RNNs, SPEL utilizes its own prior prediction errors as inputs to its hidden units.

Key Points:

• SPEL's sequence prediction error time courses closely align with neural activity patterns observed in macaque PFC.
• The model's hidden units demonstrate responses to task variables and sensitivity to stimulus probability, mirroring PFC neuron behavior.
• SPEL replicates prolonged response times to unexpected events, a behavior also observed in monkeys.

Conclusions:

• The findings suggest the PFC may construct internal models of temporal world structure via a sequence prediction error minimization learning rule.
• The SPEL model offers a unified theoretical framework for understanding lateral PFC function.