Revisiting the variable memory model of visual search

Abstract

How much memory does visual search have? A number of recent papers have explored this question from various points of view. In this paper, I propose a formal framework for comparing answers across different experimental paradigms. This framework is based on the “variable memory model” (Arani, Karwan, & Drury, 1984). This model has three parameters: Encoding probability (θ), recall probability (φ), and target identification probability (p′). The model can be used to generate cumulative distribution functions for reaction time (RT) or saccades. I compare the model to a dataset of RTs collected on a standard inefficient search for block 2s among block 5s. Assuming perfect identification (p′=1), I found that mean encoding probability was .33, and mean recall probability .71. The variable memory model provides a common metric for characterizing the behaviour of observers in different laboratories, in terms that are easy to relate to the memory literature.

Acknowledgments

I would like to thank David Fencsik, Steven Flusberg, Sarah Klieger, Melina Kunar, Evan Palmer, Ken Sobel, Adrian von Mühlenen, Jeremy Wolfe, and an anonymous reviewer for helpful comments on the manuscript, and Megan Hyle for data collection.

Notes

¹One weakness of the model is the assumption that only misses are possible, and not false alarms. However, in RT studies of search, with near-unlimited exposure durations, false alarms are quite rare.

²We assume p' = 1.0, and also that once a particular deployment is forgotten, it was not remembered again. Arani et al. (1984) claimed that model output did not change significantly if a deployment that was not recalled in one epoch was allowed to be recalled in a later epoch.