Predictive routing emerges from self-supervised stochastic neural plasticity

Predictive processing theories propose that the brain supervises itself, to build an internal model of its environment. This internal model emerges by minimizing the prediction error, the discrepancy between internally generated predictions and external sensory signals. Prior work has proposed that the neurobiological implementation of predictive processing involves neuronal oscillations in the gamma (40-100 Hz) and alpha/beta (10-30 Hz) frequency range. In current computational modeling approaches to predictive processing, there exists a trade-off between algorithmic and implementational aspects. One group of predictive processing models perform self-supervised computations but miss neurobiologic details and oscillatory neurodynamics. In the other group, biophysical models implement neural network models with maximal brain-like structure and oscillatory dynamics but require manual supervision. Here, we propose an evolutionary algorithm, the genetic stochastic delta rule (GSDR), to conduct simulations with biophysical neural networks that can inform predictive processing and other theories by linking the algorithmic and implementation levels. We first evaluate GSDR in a simplified and minimal optimization problem. Then, we simulate commonly observed neural dynamics such as firing rate and modulation of neuronal oscillations. We show that GSDR is capable of replicating oscillatory dynamics in the gamma and alpha/beta frequency bands observed through in-vivo electrophysiology, such that they emerge from synaptic plasticity. This methodology broadens the scope for biology-plausible, automated, large-scale and multi-objective simulations within computational neuroscience. With this virtuous cycle between data and models established by GSDR, we suggest that the search for the circuits underlying predictive processing can be grounded to neuronal data, improving the neurobiological basis of theories. Author summaryIn predictive processing theories, it is hypothesized that the brain creates an internal model of its environment. This is supported both by empirical and theoretical studies in neuroscience, which also emphasize the importance of neuronal oscillations in this process. Importantly, brain and cognitive disorders such as schizophrenia are associated with abnormal neuronal oscillations and impaired predictive processing. To gain further insights about these functions, it is important to build detailed biophysical models where the underlying mechanisms can be explored and understood. Current models either perform computations without oscillatory rhythms or are biologically detailed but require manual tuning, which limits an unbiased search through parameter space. To address this, we developed a self-supervised evolutionary algorithm, the genetic stochastic delta rule (GSDR). We show that GSDR automatically tunes biophysical neural networks to find the optimal parameters to generate brain-like oscillatory dynamics, which are key features of predictive processing theories. Therefore, GSDR can help us find the real brain circuits that perform predictive processing.