Abstract
Recall of visual features from working memory shows stimulus-specific variation in both bias and precision (Bae & Flombaum, 2014; Pratte et al., 2017). While a number of existing models can approximate the average distribution of recall error across target stimuli, attempts to capture the way in which error varies with the choice of target have been ad hoc. Here we extend Bays' (2014) neural resource model – whereby stimuli are encoded in the normalised spiking activity of a population of tuned neurons – to provide a principled account of these stimulus-specific effects. Following previous work (Ganguli & Simoncelli, 2014; Wei & Stocker, 2015), we allow each neuron's tuning function to vary according to the principle of efficient coding. This principle states that neural responses should be optimised with respect to the natural frequency of stimuli in the environment. For orientation stimuli this means incorporating a prior that favours cardinal over oblique orientations. While continuing to capture changes to the mean distribution of errors with set size, the resulting model accurately described stimulus-specific variations in recall error. Additionally, the efficient coding model predicts a repulsive bias away from cardinal orientations – a prediction that ought to be sensitive to changes in the environmental statistics. We subsequently tested whether shifts in the stimulus distribution influenced response bias to uniformly sampled target orientations. Across adaptation blocks we manipulated the cardinality of non-target array items by sampling from one of two bimodal distributions: a congruent distribution with peaks centred on cardinal orientations and an incongruent distribution with peaks centred on oblique orientations. Prior to adaptation observers were repulsed away from the cardinal axes. However, exposure to the incongruent distribution produced systematically increasing biases away from oblique orientations that persisted post-adaptation. This result confirms the role of prior expectation in generating stimulus-specific effects and validates our neural framework.
Meeting abstract presented at VSS 2018