Abstract
Although change detection is a leading visual short-term memory (VSTM) paradigm, it is still unclear exactly what information is stored and used during a change detection trial. The traditional view was that an item is stored in an all-or-none fashion. Recent work has shown that VSTM is better described as storing a noisy measurement of a stimulus. Here, we take this notion a step further and hypothesize that the level of uncertainty associated with a noisy stimulus memory is also stored and is taken into account during the decision. Subjects remembered the orientation of an ellipse and responded whether that orientation was different from a line segment's orientation presented after a delay. We manipulated VSTM uncertainty through ellipse elongation. The data were best fitted by a model in which the observer takes uncertainty into account, although not optimally. This result suggests that visual short-term memories are probabilistic and that change detection is a form of probabilistic inference. We then asked how a neural circuit could implement probabilistic change detection. We trained a recurrent two-layer neural network consisting of generic nonlinear units using trial-to-trial feedback. Training stimuli were encoded by a population of orientation-selective Poisson neurons. We found that given enough training trials, the network performs the task near-optimally, develops an explicit representation of the probability that a change occurred, and generalizes to new and variable stimulus reliabilities. Examining the dynamics of the network, we found that few hidden units exhibit the persistent activity long believed to be essential for the maintenance of a memory. Instead, most units are active only during a portion of the delay period, consistent with recent physiology. Together, our results suggest a new conceptualization of change detection at both the behavioral and the neural level.
Meeting abstract presented at VSS 2016