A network of Bayesian agents for reward prediction and noise tolerance.

The development of computational models capable of performing biologically relevant tasks is essential for real-world applications. This work introduces a network of single neurons modeled as independent agents, each optimizing its own cost function. This optimization governs the firing probability, resulting in spike generation-the core unit of neural communication. The network is hierarchically structured to reflect the connectivity of the dopaminergic reward system. Neurons adapt their synaptic weights to improve input prediction, leading to the self-organization of circuit activity. While the network performs well on classical reward-based tasks, its performance degrades when noise is introduced at the level of individual neurons, affecting spike generation. In line with biological systems, spike generation is inherently noisy, yet the brain achieves reliable computation through evolved mechanisms. Similarly, increasing the network size restores performance. By incorporating strategies to counteract intrinsic noise, this model lays the foundation for robust, energy-efficient, scalable, and noise-tolerant neuromorphic architectures.