Purchase this article with an account.
Rui Wang, Yuan Shen, Peter Tino, Zoe Kourtzi; Learning of hierarchical temporal structures facilitates the prediction of future events. Journal of Vision 2014;14(10):1169. doi: https://doi.org/10.1167/14.10.1169.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Previous experience is thought to facilitate our ability to extract spatiotemporal regularities from cluttered scenes. Recent work has focused on simple structures defined by associative pairing, or probabilistic sequences. However, event structures in the environment are typically hierarchical, comprising of simple repetitive to more complex probabilistic combinations (e.g. as in language, music, navigation). Here we test whether exposure to temporal sequences facilitates our ability to learn hierarchical structures and predict upcoming events. In particular, we employed variable memory length Markov models to design hierarchically structured temporal sequences of increasing complexity. We presented observers with a sequence of four different symbols that differed either in their probability of occurrence (0.2, 0.8, 0, 0) or the length of the predictive temporal context (up to a context length of two sequence items). Observers were first trained with sequences determined only by probability of occurrence and then variable context length (first context length of one item, then context length of two items). In each trial, the sequence was interrupted by a test stimulus and observers were asked to indicate whether the test symbol matched their expectation based on the preceding sequence. Our results demonstrate different learning profiles for hierarchically structured sequences across observers based on probability maximization vs. matching. Successful learners learned quickly (within 2 training sessions) to predict the most frequent symbol for each context (i.e. probability maximization). In contrast, weak learners based their predictions on symbol probabilities (i.e. probability matching) and required more (4-5) training sessions to learn the correct hierarchical structure. This predictive learning ability was compromised when symbol probabilities were less discriminable, but not when the sequence structure was less constrained. These findings suggest that learning context-dependent probabilities rather than memorizing temporal positions in hierarchical structures facilitates the prediction of upcoming events in complex environments.
Meeting abstract presented at VSS 2014
This PDF is available to Subscribers Only