Pupil diameter encodes the idiosyncratic, cognitive complexity of belief updating

Abstract

Pupils tend to dilate in response to surprising events, but it is not known whether these responses are primarily stimulus driven or instead reflect a more nuanced relationship between pupil-linked arousal systems and cognitive expectations. Using an auditory adaptive decision-making task, we show that evoked pupil diameter is more parsimoniously described as signaling violations of learned, top-down expectations than changes in low-level stimulus properties. We further show that both baseline and evoked pupil diameter is modulated by the degree to which individual subjects use these violations to update their subsequent expectations, as reflected in the complexity of their updating strategy. Together these results demonstrate a central role for idiosyncratic cognitive processing in how arousal systems respond to new inputs and, via our complexity-based analyses, offer a potential framework for understanding these effects in terms of both inference processes aimed to reduce belief uncertainty and more traditional notions of mental effort.

Movie viewing elicits rich and reliable brain state dynamics

“…PD may therefore also link, at least to some extent, to transient interoceptive mechanisms. Previous work has shown evoked PD in the absence of visual stimuli may reflect higher-order cognitive processes, such as updated sensory expectations following surprise 31 , consistent with the association between larger PD and State 1.…”

The comparable strategic flexibility of model-free and model-based learning

“…Larger values of ! "#$ correspond to models with higher information complexity (Filipowicz et al, 2020;Gilad-Bachrach et al, 2003;Glaze et al, 2018). Predictive accuracy was measured as the mutual information between the model and future observations; i.e., %&$&'( = ( ; %&$&'( ).…”

“…Given the Markovian nature of both the process generating the stimuli and the processes guiding the models that are generally used to capture subject strategies in this task, past features included only elements of the previous trial and did not extend further in the past. Similar to previous applications of this method to experimental data, we also assumed that subjects treated the task in a Markovian manner, by including the latent variable 1 *+ * in the past-feature vector as a proxy for the history of previously observed transitions (Filipowicz et al, 2020;Glaze et al, 2018). We also omitted the second-step response from the past-feature vector, because the information provided by this past element does not inform the first-step responses in simulations beyond the information provided by the combination of the second-step transition (S2) and the reward (Rw), and including this element in the past-feature vector did not improve our ability to distinguish between simulations of model-based and model-free agents.…”

“…1c). We then compared these data-driven metrics of model-free and model-based tendencies to each subject's strategic complexity using our method inspired by the information-bottleneck (Filipowicz et al, 2020;Gilad-Bachrach et al, 2003;Glaze et al, 2018;Tishby et al, 2000;Fig 2a), which also is data-driven and thus did not require any explicit assumptions about the specific strategy each subject used to perform the task. Because the main effect and interaction terms were negatively correlated with each other (Spearman's rho=-0.26, p=0.0002), we computed semi-partial Spearman correlations between complexity and each variable (main effect and interaction), while accounting for the other variable.…”

“…This form of complexity can be estimated in decision-making tasks using a method inspired by the information-bottleneck, which measures the amount of past information subjects use to make decisions (Tishby, Pereira, & Bialek, 2000). In general, strategies that incorporate more past information are regarded as being more flexible, because this increased information indicates a greater diversity in how past observations influence current behavior (Filipowicz, Glaze, Kable, & Gold, 2020;Gilad-Bachrach, Navot, & Tishby, 2003;Glaze, Filipowicz, Kable, Balasubramanian, & Gold, 2018). Thus, while model-based learning strategies may be more computationally complex, this does not guarantee that they are more information complex, and therefore does not also guarantee that they are more flexible.…”