Danlei Chen, Carol Jew, Benjamin Zinszer, and Rajeev Raizada
Distributional learning research has established that humans can track the frequencies of sequentially presented stimuli in order to infer the probabilities of upcoming events (e.g., Hasher & Zacks, 1984). We hypothesize that as people learn this frequency information, probabilistically weighted representations of the next stimulus are activated in the brain prior to each trial. We present behavioral evidence that these weighted representations are measurable in the response time of the subsequent trial, and we propose a further experiment to directly test the neural hypothesis. In the behavioral experiment, twelve adult participants viewed photographs of faces, tools, and buildings while performing a simple classification task. Each of these categories reliably evokes stronger responses in specific sets of brain areas compared to other categories (Chao & Martin, 2000; Epstein & Kanwisher, 1998; Kanwisher, McDermott, & Chun, 1997), allowing us to measure the intensity of brain activity separately and in parallel for each category in the MRI scanner. The frequency of each category (60%, 30%, 10%) was counterbalanced across six different frequency distributions. Using a two-way (Frequency-by-Category) linear mixed-effects model, we compared response times for the stimuli in each distribution to see whether the anticipation of a more frequent category reduced the response time. Response times significantly decreased with greater frequencies (t(6123) = -7.289, p < .0005), indicating that participants anticipated the stimuli proportional to the probability of the category and thereby reduced response times for the more frequent categories. With this evidence of probabilistic anticipatory representations, we are now testing this effect using functional MRI. We hypothesize that anticipation of a category will evoke activity in category-specific regions proportional to the probability of that category. If the neuroimaging results are in line with this hypothesis, they will suggest that learned distributional information produces probabilistically weighted representations of possible outcomes.