For all blocks, the image presentation threshold was the presentation duration required for a participant to achieve 75% accuracy on the task. For some participants, not all blocks yielded a stable threshold. Because of the adaptive nature of the staircasing algorithm, very poor performance at the beginning of a block could lead to a considerable number of trials at the 200-ms image duration (the ceiling duration), resulting in artificially high final threshold calculations. For all analyses reported, we excluded data from blocks in which more than 10% of trials were spent at the maximum duration of 200 ms. Altogether these trials constituted only 5% of the data, and were evenly distributed between global-property and basic-level classification, t(13) < 1. We discuss here two processing benchmarks: (a) the upper bound of the exposure duration necessary for a given categorization block (the maximum image duration for each participant during that block) and (b) the duration at which participants achieved 75%-correct performance (used to compare the time needed for equivalent performance across blocks).
To ensure that the tasks were equally difficult, we compared the maximum image exposures needed. As image presentation times were controlled adaptively in the staircase procedure, the longest presentation time for a participant corresponds to the duration at which that participant made no classification errors (recall that errors resulted in increased subsequent presentation times). If the global-property and category tasks were of comparable difficulty, we would expect them to have similar maximum durations. Indeed, the mean maximum duration was 102 ms for the global-property tasks and 97 ms for the category tasks, t(19) < 1 (see ). This result indicates that the two types of tasks were of similar difficulty.
Presentation-Time Thresholds and Maximum Image Exposures
A classic method for estimating thresholds from up-down staircases such as ours is to take the mode stimulus value for a participant (Cornsweet, 1962
; Levitt, 1971
). The logic in this experiment is simple: Because presentation duration was reduced by 30 ms following a correct response and increased by 10 ms following an incorrect response, participants converged on 75%-correct performance over the course of a block (Kaernbach, 1990
), viewing more trials around the perceptual threshold than above or below it.
As estimation with the mode is a rather coarse method, we also estimated thresholds by fitting a Weibull function to the accuracy data for each participant for each block (using the maximum likelihood procedure) and solving for the threshold. This function typically provides very good fits to psychometric data (Klein, 2001
). shows the Weibull fits and histograms of presentation times for 1 participant for a global-property block and a category block.
Fig. 3 Example of threshold computation for (a) a block of global-property classification (concealment) and (b) a block of basic-level categorization (ocean). Each graph shows data from the same participant. The top row shows accuracy as a function of presentation (more ...)
The thresholds reported in are the averages of the estimates obtained using the two methods. The presentation-time thresholds for all 14 blocks were remarkably short:
All were well under 100 ms; values ranged from 19 ms (naturalness) to 67 ms (river).
We compared the threshold values for the global-property blocks with the threshold values for the categorization blocks and found that the mean threshold (based on the average of the Weibull fit and mode value) was significantly lower for global-property classification (34 ms) than for basic-level categorization (50 ms), t(19) = -7.94, prep= .99; the difference was also significant when we used the Weibull fits only, t(19) = 7.38, prep > .99, and when we used the modes only, t(19) = 3.51, prep= .98. Note that to compare performance for different features, it is necessary to ensure that there were equivalently difficult distractor images. On the one hand, a distractor that differed from the target by only 1 pixel would produce extremely large presentation-time thresholds (if observers could perform the task at all). On the other hand, distinguishing targets from white-noise distractors should result in ceiling performance. In our tasks, distractors were always prototypically different from targets. That is, in the global-property blocks, the distractors represented the opposite pole of the queried property, and both targets and distractors came from several basic-level categories. In the categorization blocks, the distractors were prototypes of a variety of nontarget categories and were chosen so as to show the greatest variety of category prototypes. In this way, targets and distractors were chosen, to the extent possible, to vary only in the attribute being tested. Recall that ceiling performance was reached at similar presentation durations in the global-property and category tasks, which indicates that although performance had an early advantage in global-property blocks, this advantage was not due to the blocks of basic-level categorization being harder than the global-property blocks.
shows the distributions of participants’ presentation-duration thresholds for both types of tasks, using a Gaussian fit. The distributions of participants’ thresholds in categorization blocks were rather homogeneous in terms of both means and variances (see ). In contrast, the distributions of thresholds in global-property blocks () were more heterogeneous, with some coming very early and others more closely resembling the categorization thresholds. We calculated 95% confidence intervals around the means and found that the presentation-duration threshold was significantly shorter for forest than for other basic-level categories, and was significantly longer for openness and transience than for other global properties.
Fig. 4 Distributions of observers’ presentation-duration thresholds for (a) global-property classification and basic-level categorization, (b) each block of basic-level categorization, and (c) each block of global-property classification. All shown curves (more ...)