A genetic algorithm for self-supervised models of oscillatory neurodynamics

Abstract

Predictive processing theories propose that the brain builds internal models of its environment by reducing the discrepancy between internally generated predictions and external sensory signals. Prior work has linked these processes to oscillatory activity in gamma (40-100 Hz) and alpha/beta (10-30 Hz) frequency ranges. Current computational approaches face a trade-off: abstract predictive-processing models can implement self-supervised computations but often omit oscillatory spiking dynamics, whereas biophysically constrained spiking models can generate neural rhythms but often require extensive manual tuning. Here, we introduce the Genetic Stochastic Delta Rule (GSDR), an evolutionary optimization framework for fitting nonlinear neural models to electrophysiological objectives. We first evaluate GSDR in simplified optimization settings, then apply it to spiking-network objectives involving firing rates, beta/gamma spectral ratios, and empirical macaque stimulus-evoked gamma dynamics from visual cortex. We show that GSDR can search constrained synaptic parameter spaces, reduce reliance on manual tuning, and reproduce spectral and circuit-level phenotypes associated with predictive routing. We also used Izhikevich simulations as a model-class robustness analysis, showing that the approach is not limited to the original Hodgkin-Huxley-style implementation. These results position GSDR as a methodological framework for multi-objective exploration of oscillatory neural models. Author summaryIn predictive processing theories, the brain is hypothesized to build internal models of its environment. Empirical and theoretical studies suggest that neuronal oscillations are important components of this process, and abnormal oscillations are also linked to disorders such as schizophrenia. To study such mechanisms, computational neuroscience needs models that can express biologically meaningful spiking and oscillatory dynamics while also being trainable without extensive manual tuning. We developed the Genetic Stochastic Delta Rule (GSDR), a self-supervised evolutionary optimization framework for fitting nonlinear neural models to objectives. GSDR combines objective-guided search, stochastic exploration, genetic selection/deselection, and an activity-dependent MCDP update term. We show that GSDR can tune spiking networks toward beta/gamma spectral objectives and empirical stimulus-evoked gamma dynamics. The results do not prove predictive routing or identify a unique biological circuit; rather, they show that GSDR can identify candidate circuit configurations and can generalize beyond the original Hodgkin-Huxley-style model to Izhikevich simulations.