Self-organization and dynamics reduction in recurrent networks: stimulus presentation and learning.

Area of Science:

• Computational neuroscience
• Artificial neural networks
• Dynamical systems theory

Background:

• Rabbit olfactory bulb studies revealed chaotic signal dynamics linked to learned stimulus recognition via attractor dimension reduction.
• The generalizability of these findings across different neural architectures remains an open question.

Purpose of the Study:

• To investigate if chaotic dynamics and dimension reduction during learning are general properties of neural systems.
• To analyze the dynamics of random recurrent neural networks (RRNNs) and the impact of learning on their behavior.

Main Methods:

• Utilized mean-field theory to analyze the autonomous dynamics of RRNNs.
• Introduced a Hebb-like learning rule as a self-organization process.
• Numerically simulated the network dynamics under static random stimuli and during learning/recognition.

Main Results:

• Observed significant changes in the dynamical regime of RRNNs upon introduction of static random stimuli.
• Demonstrated a reduction in dynamical complexity during learning and recognition processes.
• Analyzed the dynamical repartition of local neural activity as a key mechanism.

Conclusions:

• The observed chaotic dynamics and dimension reduction during learning are not exclusive to the rabbit olfactory bulb.
• Random recurrent neural networks exhibit similar dynamical properties, suggesting general principles in neural computation.
• Self-organized learning can induce specific stimulus reactivity and simplify network dynamics.