Maheu, M. 1, 2 , Meyniel, F. 1 & Dehaene, S. 1, 3
1 NeuroSpin (CEA, INSERM), France
2 Université Paris Descartes, France
3 Collège de France, France
Our daily environment entails various temporal structures that unfold over different timescales (e.g. from the short span of a sentence to the slow cycle of seasons) and afford different predictability power (e.g. from the uncertain weather forecast to the sure sequence of traffic lights). How does the human brain manage to learn those different temporal structures? Using MEG, fMRI and finger-tracking recordings, we show that the brain is equipped with distinct systems that track different aspects of temporal sequences, that occur at different timescales. More specifically, some brain systems detect repeating deterministic patterns while some others estimate statistical trends (of various types). Importantly, we advocate two important computational properties ruling the inference of these different brain systems and how they interact with each other. First, they run in parallel based on the same input sequence. Second, they express their inference in a common probabilistic currency (thereby following normative principles). These two computational properties explain how human sensitivity to various temporal regularities emerges, but also why humans are able to quickly detect changes in their environment as well as how they manage to keep entertaining accurate predictions of the future.