The brain has a remarkable ability to infer the causes of its sensory inputs. This is clear when we consider the sensations caused by our own self-motion: when we walk or run forward, the image of the world on the retina moves towards and past us. Despite this, we do not perceive the world to be moving - we in fact perceive a stable world. On the other hand, when a train drives past, we do perceive the movement. This implies that our brain has mechanisms for distinguishing sensations caused by our own movements and those caused by the environment around us.
Figure: single neuron electrical activity was recorded as mice ran through virtual reality tunnels.
One way the brain could achieve this is to cancel out the sensory consequences of our own self-motion. This is a central feature of predictive processing models of the brain. In these models, the brain learns to predict the sensory inputs caused by our own movements during experience, and then compares these predictions to the actual sensory feedback received during the movement. The sensation caused by self-motion is cancelled-out, and all that remains is any difference between the predicted and actual sensory input, termed a ‘prediction error’.
Neurons consistent with signalling such prediction errors have been found in several regions throughout the brain. Using calcium imaging, it was previously found that in the primary visual cortex of mice, neurons in layer 2/3 respond to mismatches between mouse locomotion speed, and the speed of visual motion in a virtual reality tunnel (Keller et al., 2012). It was unclear exactly how these neurons computed such mismatches, and whether they were particularly specialized for this role.
To investigate this, we used intracellular recordings of the electrical activity of individual neurons in mice locomoting in virtual reality – a powerful method for assessing neuronal computations. We recorded activity during sudden mismatches between mouse locomotion and visual feedback, as well as during a condition in which visual flow speed is decoupled from locomotion speed. We found that neurons in layer 2/3 are particularly sensitive to mismatches between locomotion and the resulting visual feedback. Such neurons appear to make a comparison between locomotion speed and visual flow speed by subtracting one from the other, with separate sets of neurons calculating whether there is more or less visual motion than expected. Neurons in layer 2/3 seem particularly specialized for this computation, as neurons recorded deeper in the cortex, at layers 5 and 6, did not have the same properties, and likely have a different function altogether.
These results provide a key step towards comprehending how the cortex predicts the sensory consequences of our movements and accounts for them to give rise to stable perception. Understanding how the brain achieves this is important, since disruptions in this process may underly the symptoms of certain psychiatric disorders, such as schizophrenia.