Abstract
It is impossible to understand memory errors in a stimulus space without understanding the structure of that space, and taking into account a stimulus spaces’ psychophysical similarity allows straightforward noise-based models to account for nearly all aspects of memory performance (Schurgin et al. 2020). As a novel, critical test of the view that psychophysical similarity provides a unifying account of memory errors, we examined whether similarity data can be used to generalize to new domains with qualitatively distinct structures. We used a generative adversarial network to generate 3 novel “face wheels”: small, medium and large. For the small wheel, all faces on the wheel are a slight variation of a single individual; whereas on the large wheel, faces varied on a wide range of dimensions, including age, race, gender, etc. Participants (N=50) memorized either 1, 2, or 4 faces from each of these wheels and reproduced them. The wheels generated dramatically different memory error distributions: when remembering 1 item from the small wheel, a large number of participants’ errors were >90deg; whereas even when remembering 2 items from the large wheel, no errors exceeded 20deg. Can similarity structure predict these error distributions? We find that it can: independent similarity data (N=50) allowed 0-free-parameter predictions of the shape of these distributions, such that a single memory strength parameter from a given set size could predict the wildly-divergent shapes of the error distributions on the three wheels. By contrast, while alternative assumptions about the source of memory errors can sometimes also fit — though not predict — performance for simple stimuli (Tomic & Bays, 2022), these models could not fit this data. Overall, we find that a model using independent similarity measures can predict — rather than simply fit — qualitative and quantitative characteristics of memory errors in novel, realistic stimulus spaces.