Input-driven circuit reconfiguration in critical recurrent neural networks.

Area of Science:

• Computational Neuroscience
• Network Science
• Dynamical Systems

Background:

• Circuit reconfiguration, the dynamic alteration of hardware functionality, has significant technological applications.
• The cerebral cortex exhibits circuit reconfiguration, suggesting its principles may elucidate brain function.
• Understanding self-reconfiguration mechanisms is crucial for advancing both artificial and biological systems.

Purpose of the Study:

• To present a simple model of dynamical circuit reconfiguration.
• To demonstrate how input signals can dynamically alter network pathways without changing synaptic weights.
• To explore the potential of such networks in solving computational problems like connectedness detection.

Main Methods:

• Utilized single-layer convolutional recurrent networks with local unitary synaptic weights and sigmoidal activation.
• Generated traveling waves using high spatiotemporal input frequencies.
• Employed low spatiotemporal input frequencies to guide traveling waves through input-specified spatial patterns.

Main Results:

• Networks demonstrated dynamic pathway switching solely through input modulation.
• The mechanism leverages properties of marginally stable, dynamically critical systems inherent in unitary convolution kernels.
• Successfully solved the connectedness detection problem by controlling signal propagation pathways.

Conclusions:

• A simple family of networks can achieve dynamical reconfiguration through input control.
• This mechanism offers a new perspective on how biological systems, like the brain, might dynamically manage information flow.
• The demonstrated ability to solve the connectedness problem highlights the practical potential of these reconfigurable networks.