A genetic algorithm for self-supervised models of oscillatory neurodynamics
Hamed Nejat1, Jason Sherfey2, André M Bastos1,3
1Department of Psychology, Vanderbilt University, Nashville, Tennessee, United States of America.
Plos One
|August 5, 2026
Summary
We developed a new computational method, the Genetic Stochastic Delta Rule (GSDR), to better model brain activity. GSDR automates the tuning of neural network models, improving our understanding of predictive processing and neural oscillations.
Area of Science:
- Computational Neuroscience
- Systems Neuroscience
- Neural Oscillations
Background:
- Predictive processing theories suggest the brain uses internal models to minimize prediction errors from sensory input.
- Neural oscillations, particularly in gamma (40-100 Hz) and alpha/beta (10-30 Hz) bands, are linked to these predictive processes.
- Existing computational models face a dilemma: abstract models lack spiking dynamics, while biophysical models require extensive manual tuning.
Purpose of the Study:
- Introduce the Genetic Stochastic Delta Rule (GSDR), an evolutionary optimization framework for fitting nonlinear neural models.
- Address the trade-off between abstract predictive-processing models and biophysically constrained spiking models.
- Enable automated, multi-objective exploration of oscillatory neural models.
Main Methods:
- Developed the Genetic Stochastic Delta Rule (GSDR), an evolutionary optimization framework.
- Applied GSDR to fitting nonlinear neural models to electrophysiological objectives, including firing rates and spectral ratios.
- Evaluated GSDR in simplified settings and on spiking-network objectives using macaque visual cortex data.
- Performed model-class robustness analysis using Izhikevich simulations.
Main Results:
- GSDR effectively searches constrained synaptic parameter spaces, reducing manual tuning.
- The framework successfully reproduces spectral and circuit-level phenotypes associated with predictive routing.
- GSDR demonstrated robustness across different neural model implementations (e.g., Hodgkin-Huxley, Izhikevich).
Conclusions:
- GSDR provides a methodological framework for automated, multi-objective fitting of oscillatory neural models.
- This approach facilitates the study of neural dynamics underlying predictive processing.
- Predictive routing serves as a key application case, demonstrating the framework's utility.
