In the introductory post about prediction, I briefly mentioned that prediction occurs at many different levels. Even cells on the retina appear to predict stimulus motion and to alert the brain when the stimulus changes its direction unexpectedly.
In this post, I want to describe how ganglion cells extrapolate motion information and how they detect discrepancy between predicted and actual motion.
Motion extrapolation in the retina
This anticipatory firing of the retinal cells is consistent with a gain control mechanism. Following gain control, the activity of a cell is actively reduced (by negative feedback) after a period of activity. Therefore, the response to any subsequent input is dampened. A moving bar should elicit activity in a ganglion cell for an extended period of time when the bars crosses the cell's receptive field. Actually, the initial activity evoked by the bar is normal but, due to negative feedback, any subsequent activity is strongly reduced (by a factor 4!!!). In other words, peak activity of the retinal ganglion cells corresponds to the entrance of the stimulus in its receptive field and this initial burst of firing inhibits any further activity that should occur when the moving bar crosses the receptive field. The fact that peak activity corresponds to the entrance of the stimulus in the receptive field yields motion extrapolation.
In summary, the very simple mechanism of gain control can account for motion extrapolation in the retinal ganglion cells. But these cells are extraordinary as they can also detect when their prediction went wrong!!! In other words, the retina trumpets its failed prediction (Holy 2007).
Prediction error detection
When the bars moved smoothly across the receptive fields of a cell, they recorded the same firing patterns as in the 1999's study, i.e. the leading edge of the moving bar elicit a burst of spikes, which starts before it reaches the center of the receptive field. However, on some instance, the bar unexpectedly reversed its direction of motion. Although the cells responded similarly to motion in all direction, they did not exhibit the usual burst of spikes immediately after the bar reversed its direction. The activity evoked by a moving bar just after motion reversal is very different than the activity evoked by a bar that would not have reversed immediately before. Rather, 250ms later, the cells exhibit a huge peak firing rate to signal the error in the motion reversal. This burst of activity was synchronized in many cells, allowing a very simple decoder scheme to predict the occurrence of motion reversal from the activity of the cell population. In addition to signal the prediction error, this firing burst also helps the retina to catch-up with the stimulus.
In summary, although the processing of visual information takes about 50ms at the level of the retina, circuits in the retina itself are able to partially compensate for these delays by predicting future target motion. In addition, the retinal is able to signal when its own prediction failed. All in all, the retina is much smarter than what people usually believed (and especially, scientists) (Gollisch and Meister, 2010).
Berry, M. J., Brivanlou, I. H., Jordan, T. A., & Meister, M. (1999). Anticipation of moving stimuli by the retina. Nature, 398(6725), 334-8. doi: 10.1038/18678.
Schwartz, G., Taylor, S., Fisher, C., Harris, R., & Berry, M. J. (2007). Synchronized firing among retinal ganglion cells signals motion reversal. Neuron, 55(6), 958-69. doi: 10.1016/j.neuron.2007.07.042.
Holy, T. E. (2007). A public confession: the retina trumpets its failed predictions. Neuron, 55(6), 831-2. doi: 10.1016/j.neuron.2007.09.002.
Gollisch, T., & Meister, M. (2010). Eye smarter than scientists believed: neural computations in circuits of the retina. Neuron, 65(2), 150-64. Elsevier Inc. doi: 10.1016/j.neuron.2009.12.009.