Abstract
Considerable evidence from behavioral and neural studies indicates that faces are detected rapidly. However, a face reveals not just the presence of a person, but many different kinds of information about that person, such as their gender, age, familiarity and specific identity. How quickly are these specific dimensions of face information represented? To find out, we used Magnetoencephalography (MEG) and decoding techniques to measure the time course of extraction of each of these dimensions of face information. Subjects (n = 5) performed a one-back task while viewing 80 different face images; each image was repeated 25 times. The stimulus set consisted of five different images from each of 16 different celebrities. We chose the celebrities such that half of them were familiar (US actors) versus unfamiliar (German actors), young (< 36 years) versus old (> 59 years), and female versus male. MEG decoding accuracy was computed separately at each time point (10 ms bins) after stimulus onset based on stimuli that varied in lighting, pose and orientation (for face identity decoding), and across individuals (for familiarity, gender and age decoding). In each subject individually, we could decode all dimensions of face information within the first 150 ms after stimulus onset (mean peak decoding accuracy about 70% for binary dimensions, and about 30% for 16-way identity). This decoding of face dimensions occurs at shorter latencies than position and size invariant object category decoding (Isik et al 2014). Further, these results cannot be easily explained by low-level differences, because all face dimensions could be decoded rapidly, whether they contained discriminative low-level features (e.g. age) or not (e.g. familiarity) as measured by decodability from early layers of a face deep neural network (VGG-Face). Overall, our results indicate that many different dimensions of face information become available extremely rapidly, within 150 ms of stimulus onset.
Meeting abstract presented at VSS 2018