Motor primitive and sequence self-organization in a hierarchical recurrent neural network

Area of Science:

• Computational neuroscience
• Robotics
• Artificial intelligence

Background:

• Hierarchically organized recurrent neural networks (RNNs) offer a framework for understanding complex behaviors.
• Adaptation processes are crucial for developing sophisticated goal-directed actions.
• Simulated environments provide a controlled setting for testing artificial intelligence models.

Purpose of the Study:

• To investigate how complex goal-directed behavior can arise from adaptation in a hierarchical RNN.
• To explore the self-organization of dynamic structures within the network for navigation tasks.
• To analyze the underlying mechanisms of behavior primitive switching and motor sequence generation.

Main Methods:

• Utilized a genetic algorithm (GA) to train a hierarchically organized recurrent neural network.
• Employed a simulated Khepera robot for experiments in complex navigation tasks.
• Analyzed emergent dynamic structures at different network levels (lower and higher).

Main Results:

• Observed self-organization of distinct dynamic structures in lower and higher network levels.
• Identified parametric bifurcation structures in the lower level, explaining top-down behavior primitive switching.
• Discovered a topologically ordered mapping in the higher level, linking activation states to motor sequences via non-linear dynamics.

Conclusions:

• Hierarchical RNNs, optimized by GAs, can self-organize complex behaviors for navigation.
• The model demonstrates biologically plausible principles for generating adaptive, goal-directed actions.
• Emergent network dynamics provide insights into the mechanisms of behavior control and learning.