In this opinion article, I describe how recent findings showing that the locus coeruleus responds to multiple types of prediction error support models in which the locus coeruleus computes global model failures: incidences where predictions about the world are strongly violated. I explain how this computation would be useful in setting the rate of learning about predictions, with reference to the cortical circuit, to balance stability and flexibility of internal models. I also explain how this view can account for other known functions of the locus coeruleus, and could be computed with known locus coeruleus circuitry.

The locus coeruleus is a neuromodulatory system that is known for responding to unexpected stimuli. One possibility is that these responses reflect prediction errors. Using two-photon calcium imaging while mice run through a virtual reality, we find that the locus coeruleus broadcasts unsigned visuomotor prediction errors in the cortex. Then, using optogenetic stimulation, we show that such transient locus coeruleus activity can function to gate cortical sensorimotor plasticity over the course of minutes, in a manner that recapitulates the type of plasticity seen over days of visuomotor development.

We review the evidence supporting the idea that visuomotor mismatch responses in V1 result from prediction error computation. We go on to show that the well known locomotion induced gain increase of visual responses in V1 cannot account for mismatch responses, because the apparent gain is strongly dependent on recent experience of visuomotor coupling. This supports the idea of a visuomotor prediction error computation in V1 that is dynamically dependent on the predictability of visual input.

Whole cell recording in vivo, particularly in moving animals, is a notoriously tricky technique to master, and this is not helped by the lack of in depth protocols detailing all the tricks of the trade. Rebecca wrote this protocol so that hopefully anyone with a whole cell recording rig and enough time to master the method has as much of the practical information as possible to achieve success, and acquire these beautiful, information rich recordings.

Prediction error computation requires a neuron to calculate the difference between a prediction of sensory input, and the sensory input actually received. In this paper, we use whole cell recordings from primary visual cortex as mice locomote in a virtual reality system. We show that individual neurons in layer 2/3 have membrane potential dynamics consistent with a comparison of visual flow input and a prediction of visual flow input based on locomotion speed. In deeper layers, this kind of computation was not apparent.