Spatiotemporal Dynamics and Reliable Computations in Recurrent Spiking Neural Networks

Area of Science:

• Computational neuroscience
• Neural network modeling

Background:

• Randomly connected spiking neural networks model neural variability but struggle with computational reliability.
• The inherent unreliability limits their application in complex computational tasks.

Purpose of the Study:

• To investigate if incorporating distance-dependent connection probabilities can improve the computational performance of spiking neural networks.
• To explore the potential of spatially extended networks for dynamical computations.

Main Methods:

• Modeling randomly connected networks of excitatory and inhibitory spiking neurons.
• Introducing connection probability dependence on the physical distance between neurons.
• Analyzing network behavior using concepts like symmetry-breaking bifurcations.
• Applying a reservoir computing framework for training and evaluating computational tasks.

Main Results:

• Spatially extended spiking networks with distance-dependent connections overcome the unreliability of random networks.
• These networks exhibit symmetry-breaking bifurcations, leading to the generation of rich spatiotemporal patterns.
• The generated patterns can be effectively trained to perform dynamical computations within a reservoir computing framework.

Conclusions:

• Distance-dependent connectivity is a crucial factor for enhancing the computational capabilities of spiking neural networks.
• Spatially organized neural networks offer a promising framework for robust and reliable dynamical computations.
• This approach bridges the gap between biologically plausible neural models and practical computational applications.