I discuss how recent findings indicate that the locus coeruleus responds to various prediction errors, supporting models where the LC signals major failures in prediction. I illustrate how this function is useful in adjusting the learning rate of cortical predictions, helping to balance the stability and flexibility of internal models. Additionally, I show how this perspective aligns with other established functions of the locus coeruleus and can be processed using its known circuitry.

The neuromodulatory locus coeruleus (LC) responds to unexpected stimuli, potentially reflecting prediction errors. Through two-photon calcium imaging of mice in virtual reality, we discovered that the LC signals unsigned visuomotor prediction errors to the cortex. Optogenetic stimulation reveals that transient LC activity can gate cortical sensorimotor plasticity within minutes, mimicking the plasticity usually observed over days of visuomotor experience.

​

We review the evidence that visuomotor mismatch responses in V1 arise from prediction error computation. We demonstrate that the locomotion-induced gain increase in V1 visual responses does not explain mismatch responses, as it heavily relies on recent visuomotor coupling experience. This supports the notion of a dynamic visuomotor prediction error computation in V1 based on visual input predictability.

Whole cell recording in vivo is especially challenging in moving animals, in part due to the scarcity of detailed protocols. Rebecca created this protocol to provide essential practical information for anyone with a whole cell recording rig to successfully master the method and obtain valuable recordings.

Prediction error computation involves a neuron computing the difference between predicted and actual sensory input. This study uses whole recordings in the primary visual cortex of mice running in a virtual reality system, revealing that layer 2/3 neurons exhibit membrane potential dynamics indicative of comparing visual flow with predictions based on locomotion speed, while deeper layers do not show this computation.