| name | multi-objective-optimisation-oscillatory-snn |
| description | Multi-objective genetic algorithm (NSGA-III) optimisation of Izhikevich neuron-based recurrent spiking neural networks for simultaneously matching neural firing rates and network oscillation frequencies. Based on arXiv:2605.25224 (May 2026). Use when studying SNN parameter fitting, neural oscillations, genetic algorithm optimisation for spiking networks, or brain organoid modeling. |
Multi-Objective Optimisation with Oscillatory Dynamics in Spontaneous and Decision Spiking Neural Networks
Methodology from arXiv:2605.25224 (May 2026).
Authors: Divyansh Sethi, Muhammad Faraz, KongFatt Wong-Lin
Subjects: q-bio.NC
Overview
Spiking neural networks (SNNs) can be used for cost-efficient AI computing or mechanistic modeling of neural data. Fitting recurrent SNNs (RSNNs) to neural data remains challenging — especially when simultaneously matching both neural firing rates and network oscillation frequencies, both of which are known to play important roles in neural function.
This work extends the application of NSGA-III (Non-dominated Sorting Genetic Algorithm III) to Izhikevich neuron-based RSNNs by optimizing connectivity parameters to target emergent neuronal (sub)population firing rates AND network oscillation frequencies simultaneously.
Key Contributions
- Multi-objective GA for RSNN parameter fitting: Uses NSGA-III to simultaneously optimize for firing rates and oscillation frequencies
- Validation on three regimes: Spontaneously active RSNN, low-activation brain organoid, and simulated decision-making RSNN
- Parameter sensitivity analysis: Found that dominant oscillation frequencies are more parameter-sensitive than firing rates
- Low-activity regime identification: Identified distinct low-activity regime for decision-making dynamics
Methodology
Network Architecture
- Izhikevich neuron model for cortical excitatory and inhibitory neurons
- Recurrent connectivity with spontaneous firing dynamics
- Models comprise spontaneously firing cortical excitatory and inhibitory populations
Optimization Framework
- Algorithm: NSGA-III (multi-objective genetic algorithm)
- Objectives: Minimize RMSE between target and emergent:
- (Sub)population firing rates
- Dominant network oscillation frequencies
- Decision variables: Connectivity parameters of the RSNN
- Evaluation: Pareto frontier analysis
Three Validation Scenarios
| Scenario | Description | Key Finding |
|---|
| Spontaneous RSNN | Simulated spontaneously active RSNN | Both targets can be simultaneously optimized |
| Brain Organoid | Low-activation brain organoid model | Multi-objective optimization applicable to biological systems |
| Decision-Making RSNN | Transient decision dynamics with temporal epochs | Activity patterns in different time epochs successfully fitted |
Key Findings
- Oscillations are more sensitive: Dominant oscillation frequencies are harder to fit than firing rates, showing higher parameter sensitivity
- Firing rates are more robust: Firing rate targets are more reliably achieved across different parameter settings
- Low-activity decision regime: Identified distinct low-activity dynamical regime in decision-making networks
- Pareto frontier trade-offs: RMSE-based Pareto frontier enables analysis of multi-objective trade-offs
Practical Implications
- Neural data fitting: Provides methodology for fitting RSNNs to experimental neural recordings with both rate and oscillation constraints
- Brain organoid modeling: Demonstrates applicability to biological neural systems ex vivo
- Neuromorphic computing: Multi-objective optimization could improve SNN-based AI systems
- Clinical applications: Understanding parameter sensitivity can inform brain disorder modeling
Activation Keywords
- spiking neural network, NSGA-III, multi-objective optimization
- neural oscillations, Izhikevich neuron, recurrent SNN
- brain organoid, neural data fitting, Pareto frontier
- firing rate optimization, oscillation frequency, decision-making dynamics
