Abstract
Classification images (CIs) can reveal observers' strategies in a variety of visual tasks. However, one weakness of the CI method is that many trials are needed to obtain stable data (e.g., 10,000 trials for 128 × 128 face images in Sekuer etal., 2004). We examined whether CIs can be obtained with fewer trials, thereby making it possible to use the method with clinical populations. Because the number of required trials roughly increases with the number of independent pixels, we reduced the number of pixels in our stimuli while maintaining overall stimulus size. We used the same upright faces and task used by Sekuler et al. (2004). However, instead of presenting the entire face, one pixel in each 2 × 2 region was randomly sampled, and the remaining pixels in that region were set to zero contrast. With these sampled faces, we obtained CIs from six typical and two autistic observers. After 1,450 trials, sums of squares (SS) were calculated from each CI by squaring each value and summing across the entire face. To examine the structure in more detail, CIs were filtered with a 10×10 convolution kernel, and SS values were calculated for seven regions: the forehead, nose, mouth, and the left and right eyes and cheeks. Permutation tests showed that SS values from the left and/or right eyes were significantly greater than chance levels for all observers. Additionally, the SS value from the forehead was statistically significant in one normal and one autistic observer. In conclusion, we successfully obtained face CIs for normal and autistic observers in relatively few trials. Moreover, the results with sampled faces are qualitatively similar to those obtained with normal faces. We currently are testing more autistic observers to determine how face-processing is affected by autism.
This study is supported by the CIHR-JSPS Japan-Canada Joint Health Research Program and the Canada Research Chair programme.