Abstract
Statistical regularities in the occurrence of environmental properties are useful to consider in many tasks. The representation of such regularities by the nervous system can be modeled as a prior distribution. Such a prior could represent all past experience with that property. For changing environments, however, it may be useful to inculcate recent information into the prior. One factor that influences the rate of change of a prior is the time over which the nervous system considers past information while tracking environmental changes. We performed a study to obtain an estimate of the state of the prior in a task for which the statistics of the environment change unexpectedly. Participants reached across a tablet to hit a target that moved along a fixed path to a specified location in space. The time the target took to reach the specified point of interception was normally distributed within blocks of 83 trials. The mean time differed between blocks (660 or 885 ms) but the standard deviations were identical (73.5 ms). While the target was visible in most trials, in some trials it disappeared 72 ms after its appearance. Participants were instructed that they may not see the target in some trials but they should nevertheless attempt to intercept it as if it were moving invisibly. We treat the participants' response times on such catch trials as an estimate of their prior. We model the response times of the catch trials using a Kalman filter with a time-window that recalculates the gain as the experiment progresses. We compare the squared residuals to those obtained from a model that uses an optimal fixed gain (Kalman, 1961, Burge et al., 2008). The Akaike's information criteria indicate that the model with an optimal window is three times more likely to be correct.
The research leading to these results has received funding from the European Community's Seventh Framework Programme FP7/2007-2013 under grant agreement number 214728-2.