Dissociating dynamic probability and predictability in observed actions-an fMRI study

Abstract

The present fMRI study investigated whether human observers spontaneously exploit the statistical structure underlying continuous action sequences. In particular, we tested whether two different statistical properties can be distinguished with regard to their neural correlates: an action step's predictability and its probability. To assess these properties we used measures from information theory. Predictability of action steps was operationalized by its inverse, conditional entropy, which combines the number of possible action steps with their respective probabilities. Probability of action steps was assessed using conditional surprisal, which increases with decreasing probability. Participants were trained in an action observation paradigm with video clips showing sequences of 9-33 s length with varying numbers of action steps that were statistically structured according to a Markov chain. Behavioral tests revealed that participants implicitly learned this statistical structure, showing that humans are sensitive toward these probabilistic regularities. Surprisal (lower probability) enhanced the BOLD signal in the anterior intraparietal sulcus. In contrast, high conditional entropy, i.e., low predictability, was correlated with higher activity in dorsomedial prefrontal cortex, orbitofrontal gyrus, and posterior intraparietal sulcus. Furthermore, we found a correlation between the anterior hippocampus' response to conditional entropy with the extent of learning, such that the more participants had learnt the structure, the greater the magnitude of hippocampus activation in response to conditional entropy. Findings show that two aspects of predictions can be dissociated: an action's predictability is reflected in a top-down modulation of attentional focus, evident in increased fronto-parietal activation. In contrast, an action's probability depends on the identity of the stimulus itself, resulting in bottom-up driven processing costs in the parietal cortex.

Keywords: action observation; dmPFC; fMRI; information theory; orbitofrontal cortex; statistical learning.

Figure 1

Markov chain ruling the presented action sequences. Rows depict the first objects of a transition (t - 1), e.g., the board (first row) was always (p = 1.0) followed by a cube (third column), whereas the cube (third row) could be followed by a washer (p = 0.25), a short screw (p = 0.50), or a screw nut (p = 0.25). Conditional surprisal of an action step depended on its probability given the preceding action step only. An example is highlighted in the figure: cells surrounded by dotted lines determine the surprisal assigned to the washer after a screw nut (orange) or a cube (blue). In contrast, an action step's conditional entropy depended on its own probability and the probability weights of alternative action steps. For instance, cells surrounded by dashed lines determine the conditional entropy assigned to the washer after the screw nut (orange) or cube (blue).

Figure 2

Experimental course and exemplary distribution of conditional entropy and conditional surprisal during an action sequence. A fixation circle announced each video and 48% of the videos were followed by a two alternative forced choice question. Feedback on correctness of responses was only given during the training sessions. Within each action sequence, values of conditional entropy and conditional surprisal were ascribed to the action steps, depending on the preceding action step.

Figure 3

Results of the two post-tests. As one cross in the paper–pencil post-test corresponded to 12.5% in the computer post-test, results of the paper–pencil post-test were multiplied with the factor 12.5, to make participants' probability judgments in the two post-tests more comparable. Error bars display ± 1 SD.

Figure 4

Areas showing a positive correlation with (A) conditional surprisal and (B) conditional entropy. ant dIns, anterior dorsal insula; dmPFC, dorsomedial prefrontal cortex; IFG, inferior frontal gyrus; lOFC, lateral orbitofrontal cortex; pIPS, posterior intraparietal sulcus; PMv, ventral premotor cortex; aIPS, anterior parietal sulcus.