Learning, planning, and control in a monolithic neural event inference architecture

Area of Science:

• Computational Neuroscience
• Machine Learning
• Dynamical Systems Theory

Background:

• Understanding and modeling complex dynamical systems requires inferring hidden states and predicting future behavior.
• Sensorimotor experiences provide crucial data for learning and adapting control strategies.

Purpose of the Study:

• Introduce REPRISE (REtrospective and Prospective Inference SchEme) for learning temporal event-predictive models.
• Enable systems to infer unobservable states and optimize motor activities for goal-directed control.

Main Methods:

• Implementation using a recurrent neural network (RNN) augmented with contextual neurons.
• Learning temporal forward models of sensorimotor contingencies from simulated dynamic vehicles.
• Analyzing sensorimotor error signals to adapt neural activities and connection weights.

Main Results:

• REPRISE concurrently separates and approximates sensorimotor dynamics.
• Learned models are exploited for goal-directed, model-predictive control (approximate active inference).
• Development of event-predictive neural encodings for adaptive sensorimotor control.

Conclusions:

• REPRISE effectively models dynamical systems by integrating retrospective analysis and prospective optimization.
• The scheme facilitates adaptive, goal-directed sensorimotor control through learned predictive models.
• Contextual neurons and error signal adaptation are key to REPRISE's performance.