Abstract
When asked to remember the features of an object, do we retain single point-estimates in memory, or do we retain richer representations, such as probability distributions over feature space? Continuous color report tasks, which require participants to report the exact color of a remembered object, yield a wide distribution of errors. However, it is unclear whether this distribution of errors reflects random errors in point estimates (which would form a distribution when aggregated across trials), or if it reflects the fact that individual memories are probabilistic in nature. To distinguish between these possibilities, we developed a novel paradigm: Participants were shown a set of colorful circles and, after a brief memory delay, asked about the color of a randomly tested item. Rather than reporting a single color, participants placed 10 bets in the color space (each bet comprised 1/18 of the space). Bets could be stacked or moved to a new portion of color space. Points were awarded equal to the number of bets containing the correct color. We discouraged stacking bets by diminishing the value of higher stack levels. We found that participants have considerably richer representations than simple point estimates. The dispersion of bets was correlated with first response accuracy (the less accurate their first response, the more widely they distributed their bets), indicating that participants' internal confidence was guiding bet placement. Furthermore the distribution of bets on a trial was similar in shape to the probability distributions observed across trials, suggesting that an uncertainty distribution may be coded within individual memory representations. Finally, we found that memory performance improved, relative to the first response, when averaging across multiple bets. This performance benefit increased monotonically with each additional bet. Overall, the betting task revealed that memories are more than noisy point estimates, but are surprisingly rich and probabilistic.
Meeting abstract presented at VSS 2015