Constructing and Forgetting Temporal Context in the Human Cerebral Cortex

How does information from seconds earlier affect neocortical responses to new input? Here, we used empirical measurements and computational modeling to study the integration and forgetting of prior information. We found that when two groups of participants heard the same sentence in a narrative, preceded by different contexts, the neural responses of each group were initially different, but gradually fell into alignment. We observed a hierarchical gradient: sensory cortices aligned most quickly, followed by mid-level regions, while higher-order cortical regions aligned last. In some higher order regions, responses to the same sentence took more than 10 seconds to align. What kinds of computations can explain this hierarchical organization of contextual alignment? Passive linear integration models predict that regions which are slower to integrate new information should also be slower to forget old information. However, we found that higher order regions could rapidly forget prior context. The data were better captured by a model composed of hierarchical autoencoders in time (HAT). In HAT, cortical regions maintain a temporal context representation which is actively integrated with input at each moment, and this integration is gated by prediction error. These data and models suggest that sequences of information are combined throughout the cortical hierarchy using an active and gated integration process.