Abstract
The question whether visual short-term memory (VSTM) consists of a fixed number of discrete chunks or of a continuous resource has been intensely debated. Recent work suggests that only the former model can explain human performance, in particular the observed increase in guessing with set size. However, continuous-resource models have not taken into account that the amount of resource per item may vary across items and trials, for example due to attentional fluctuations. We developed a new model that incorporates this variability. In this model, the encoding precision of a stimulus decreases on average inversely proportionally to set size but is variable around this average. We tested this variable-precision model against leading previous models in three experimental paradigms: delayed estimation (recall), change detection, and change localization. We performed each experiment for both orientation and color. In the change detection and change localization experiments, we varied not only set size, but also the magnitude of change and stimulus reliability. Across all six experiments, the variable-precision model provided excellent fits to the data, accounted for the increase in guessing, and outperformed the other models in a formal comparison. Furthermore, we found that the variable-precision model can readily be implemented using an existing neural architecture, namely the divisive normalization model of attention (Reynolds and Heeger). We modified this model to incorporate attentional fluctuations and stochasticity in spike generation. With these modifications, the neural network can quantitatively reproduce subjects’ behavior in delayed estimation, despite not containing any limit on the number of items kept in memory. The neural model is consistent with extant physiological findings. Together, these results provide strong converging evidence – both behavioral and neural – that VSTM resource might not be discrete and fixed, but continuous and variable. The variable-precision model might provide a unifying framework of attention and working memory.
Meeting abstract presented at VSS 2012