Jarzebowski, P.; Bendor, D.

The brain refines its predictions of the world by updating its internal model whenever sensory input differs from expectation. The sign of this prediction error matters: an unexpected event signals that the model under-predicted (positive error), while a predicted event that fails to occur indicates that the model over-predicted (negative error), and the two should drive opposite synaptic changes. How cortical circuits represent error sign in spiking activity, and how that representation translates into synaptic learning, remain unresolved. We propose the Signed Error by Timing Asymmetry (SETA) model, in which the sign of a prediction error is encoded by when layer 2/3 neurons fire relative to a brief plasticity window in their layer 5 targets. Chandelier cells, an inhibitory cell type recruited by the prediction, impose a temporal clamp on layer 2/3 output: positive errors escape the clamp and arrive within the synaptic potentiation window, while negative errors are released only after the clamp decays and arrive later, during the synaptic depression window. The same circuit, therefore, biases downstream synapses toward either potentiation or depression depending on the prediction-error sign. We demonstrate this signed-error computation in a reduced two-compartment model, test SETA-specific predictions using in vivo recordings from mouse visual cortex, and examine how E/I imbalance leads to pathological consequences in predictive coding.