Intrinsic adaptation in autonomous recurrent neural networks

Area of Science:

• Computational neuroscience
• Complex systems theory

Background:

• Recurrent neural networks (RNNs) exhibit autonomous activity crucial for information processing.
• Default neural activity patterns (regular, synchronized, bursting, chaotic) influence stimulus response.
• Nonsynaptic plasticity's role in shaping these default dynamics is not fully understood.

Purpose of the Study:

• To investigate how nonsynaptic plasticity affects the default dynamical states of RNNs.
• To explore the impact of intrinsic parameter adaptation driven by information entropy optimization.
• To identify emergent dynamical regimes under these adaptation processes.

Main Methods:

• Simulated massively recurrent neural networks with intrinsic adaptation mechanisms.
• Nonsynaptic plasticity modeled as adaptation of intrinsic neural parameters (threshold, gain).
• Adaptation driven by information entropy optimization.

Main Results:

• Three distinct, globally attracting dynamical regimes were observed: regular synchronized, chaotic, and intermittent bursting.
• The intermittent bursting regime features periods of stimulus-insensitive regular activity interspersed with stimulus-sensitive chaotic bursts.
• Nonsynaptic plasticity significantly influences the emergence and stability of these regimes.

Conclusions:

• Intrinsic adaptation via nonsynaptic plasticity can self-organize RNNs into complex dynamical states.
• The intermittent bursting regime offers a potential mechanism for flexible information processing in neural systems.
• Findings align with theories of self-organized criticality and efficient brain dynamics.