Key role of neuronal diversity in structured reservoir computing

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Complex systems

Background:

• Reservoir computing models, using recurrent neural networks, capture chaotic time series.
• Existing models typically use randomly connected, homogeneous neural units.
• A need exists for structured models incorporating neural diversity and connectivity.

Purpose of the Study:

• To introduce a novel class of structured reservoir computing models.
• To investigate the impact of diverse cell types and their connections on model dynamics and performance.
• To explore how neural microcircuits balance performance and stability in time series representation.

Main Methods:

• Developed structured reservoir models with distinct cell types (pyramidal, parvalbumin, somatostatin).
• Analyzed model stability, identifying inhibition-stabilized network (ISN) and non-ISN regimes.
• Extended models to leaky integrate-and-fire neurons, preserving network architecture.
• Trained ISN and non-ISN networks to relay and generate a chaotic Lorenz attractor.

Main Results:

• Structured models demonstrated improved performance in capturing chaotic time series.
• Two distinct dynamical regimes were identified: ISN (balanced excitation-inhibition) and non-ISN (weak excitation).
• ISN networks, while high-performing, operated near stability limits, sensitive to perturbations.

Conclusions:

• Structured reservoir computing offers a framework to explore neural microcircuit function.
• Distinct dynamical regimes (ISN vs. non-ISN) impact the balance between performance and stability.
• This approach provides insights into how neural networks represent complex time series data.