Abstract
An important role of computational models is the development of testable predictions. Here, we summarize results from testing 6 predictions concerning the Binding Pool, a computational model of visual working memory, in Swan and Wyble (2014). Prediction 1: Forgetting an object from memory should improve the precision of other stored objects. This was tested by measuring memory precision when participants were cued to forget one item during a retention interval (see also Williams et al., 2013). Prediction 2: Memory precision should decline as more features are encoded. We tested this prediction by changing the number of features reported from a single object. Participants were less precise when reporting more features. Prediction 3: Reports of features from repetitions of objects should be less precise than from a single object. Contrary to the prediction, when testing recall we found similar precision values for three repetitions as a single color. Prediction 4 pertained to confidence for repeated items and was similarly not matched by the data. Prediction 5: Two similar features should bias precision of a third feature more than dissimilar features. In an experiment, participants reported the luminance of one square while we manipulated the luminance of two other squares in the memory set. The data did not support the prediction since bias was stronger for the dissimilar features. Prediction 6: Swap errors should produce high confidence responses. We validated this prediction by looking at high confident responses with high error and found that the corresponding responses were best fit by the non-target distribution instead of a uniform distribution. The failed predictions pertain to retrieval of repeated values and thus highlight the model's inability to use ensembles to compress repeated information. Future work will explore how ensemble representations should be incorporated to allow the model to represent repeated items more efficiently.
Meeting abstract presented at VSS 2016