Abstract
In visual categorization tasks, an observer's behavior depends both on the task and on the information encoded from their visual field. Where, when and how does the brain's dynamic encoding of visual information combine with task demands to become a representation of task-relevant information that supports behavior? Eight observers categorized the expression (7-AFC, 'happy,' 'surprise,' 'fear,' 'disgust,' 'anger,' 'sad' plus 'neutral') of stimuli while we measured their single-trial MEG responses. On each of the 21,000 experimental trials/observer, with Bubbles (Gosselin & Schyns, 2001) we randomly sampled pixels from the original faces using Gaussian apertures distributed across 5 one-octave spatial frequency bands (S1-B). Using mutual information (Ince et al., 2015, 2016) we quantified in each observer the dynamic coding of specific stimulus features through the whole brain—i.e. 12,773 voxels, 0-400 ms post stimulus, 4 ms resolution. To relate feature coding in the brain to the observer's categorization behavior (i.e. correct vs. incorrect), we introduce information theoretic redundancy, the 3-way interaction between MEG source time course, stimulus feature, and observer response (Ince et al., 2016). Technically, redundancy quantifies how much of the trial-by-trial stimulus variation commonly affects both the MEG signal and the observer's behavioral response (S1-A). A compelling outcome is a space x time processing pathway in the brain of the specific features that supports categorization behavior. Across observers, we show that redundancy increases in early visual cortex from the onset of visual coding, suggesting that task demands inform the uptake of task-relevant features early on. Also, redundancy shows a sharp increase at 148 ms post stimulus, midway down the left and right fusiform gyrus (S1-C), relative to the linear increase in visual face coding. This suggests a time window and a locus for the shift from generic face feature coding to specific task-relevant feature coding in the brain.
Meeting abstract presented at VSS 2017