Abstract
How early does the brain decode object categories? The issue remains controversial because dynamic brain measurements are susceptible to signal drift, filtering distortions and statistical issues. Here, using single-trial distributions and a new statistical framework building on robust statistics, we present evidence that categorical differentiation occurs in the brain at 90-110 ms following stimulus onset (median onsets). Results were reliable across testing sessions, with effects starting at mid-line or lateral electrodes, depending on subjects. The results challenge findings of very early face effect <80 ms and confirm categorical sensitivity occurs at approximately 100 ms. Eight observers categorized pictures of faces, houses, and noise textures, presented for 53 ms (n=1000 trials). Seven observers were tested twice. We filtered EEG data using 2 Hz high-pass causal Butterworth filters. Causal filters drastically remove drifts that may affect onsets, but without introducing spreading of the signal in time caused by non-causal filtering. Across categories, single-trial ERP distributions could in principle differ in many ways (e.g. location, dispersion, skewness). Using classic GLMs, we found differences corresponding to changes in onset means. Kernel density estimates and shift functions revealed uniform shifts in single-trial distributions. We applied techniques sensitive to other distributional differences (i.e. mutual information, Kolmogorov-Smirnoff, and multivariate logistic regression) and confirmed that our GLM approach did not miss earlier onsets – suggesting that massively univariate GLMs are appropriate to study categorical ERP onsets. Finally, we describe non-parametric measures of ERP difference effect sizes that are comparable across studies. Although effect sizes were unsurprisingly weak at onsets, the differences were associated with strong evidence against the null hypothesis (Bayes factors > 3). Onsets of categorical differentiation are critical to circumscribe the timing of visual processes in the brain. For the first time we provide a statistical framework to unequivocally ascertain when such differentiation occurs.
Meeting abstract presented at VSS 2013