Recurrent neural network models reveal unified mechanisms generating event-related potentials from MMN to P300
The brain must continuously extract salient cues from an immense sensory stream and represent them across regions. Multiregional responses to salient stimuli, notably mismatch negativity (MMN) in sensory and P300 in frontal areas, are among the best-characterized signals in human electrophysiology, yet the mechanisms relating them are unclear, and whether one mechanism accounts for both is unknown. We develop a hierarchical, multiregional recurrent neural network of sensory and frontal cortex to test whether short-term synaptic depression (STD) can explain both. In sensory regions, STD produces stimulus-specific adaptation, generating MMN-like responses; propagation to frontal regions produces the amplification and delay characteristic of the P300, while enhancing noise robustness, with rare stimuli occupying expanded regions of neural state space. When STD acts on frontal feedback predicting upcoming stimuli, it further amplifies responses to unexpected stimuli, suggesting attention sharpens predictions. Our results bridge synaptic-scale mechanisms and brain-wide representations, with relevance to precision psychiatry.