Timescales of Rational Inattention

Abstract

To reconcile previously reported differences in information-seeking behaviour displayed by neurodivergent children and their neurotypical peers, Jones et al. (2024) present a compelling new theory of rational information-seeking behaviour. The central premise of rational inattention is that optimal information-seeking behaviour may centre around learning to disengage from imprecise sources, that is, sources where information gain is expected to be low and imprecise. Thus, children with dyslexia may, over time, selectively disengage from further processing of text, which past experience has taught them provides minimal information gain. Neurotypical children who have experienced greater information gain from text sources, however, may persevere and attend to such sources proportionately in the future.

Rational inattention has critical bridges to previous rational accounts of curiosity, where the rational agent aims to maximise information gain by attending to sources that maximally increase the usefulness of their knowledge (e.g., Dubey and Griffiths 2020). These links to previous accounts of curiosity are worth considering in greater detail towards a unified account of information-seeking behaviour. In particular, in reconciling these accounts, one question worth pursuing is the relationship between rational attention—as characterised by rational accounts of curiosity—and rational inattention. Is rational attention the flip side of rational inattention or are different factors likely to impact the sources children choose to attend to in the search for information gain and the sources children disengage from over time. My suspicion is that at least a partial answer to this question may lie in paying closer attention (pun unintended) to two timescales of rational inattention, in particular, the real-time and the developmental timescale of rational inattention (cf. McMurray 2016).

Let us first consider the timescale of real-time processing in the proposed model. The model suggests that some reconstruction of the input is produced regardless of the precision of this reconstruction in the initial stages of processing (see Figure 1). Thus, even input that is later evaluated as imprecise is included in the early processing stages. The error between the input and the imprecise reconstruction is downweighted, retrospectively, based on the learner's experience with such stimuli (e.g., see Schütte, Mani, and Behne 2020, for evidence of such retrospective selective learning in young children). This is where rational inattention comes in. Indeed, the early reconstruction of all input is the reason why a clamped model—without such retrospective downweighting—shows improved learning from even imprecise input. Here, rational attention and rational inattention may be seen as flip sides of the same coin with the input being retrospectively weighted in essentially the same way to allow attention to sources that ensure information gain and away from sources that are unlikely to minimise uncertainty. Such an interpretation would allow for a unified framework for information seeking, bringing together the benefits of rational accounts of curiosity—in terms of estimating the future usefulness of encountered information—and the neurodivergent perspective offered by rational inattention.

The scenario with regard to the developmental timescale is, however, less clear. In particular, minimising epistemic uncertainty requires some evaluation of the learner's current state of knowledge or abilities, that is, identifying the sources that are likely to be imprecise for specific learners. Jones et al. (2024) note that formal metacognitive awareness of the imprecision of specific sources for specific learners is not a prerequisite for the model's behaviour. Although this makes sense when considering rational (in)attention in young infants and children lacking such metacognitive awareness, how does rational inattention play out across development? On the one hand, this raises questions regarding the cognitive correlate of global error—the sum of the weighted error signals (MSE) between the input and the reconstruction from different sources—especially in young infants with limited metacognitive abilities. Moreover, how would growing metacognitive awareness in early childhood factor into rational models of information-seeking behaviour? Whilst research has long suggested that young children lack extensive metacognitive awareness of their knowledge gaps (Goupil and Kouider 2019), more recent work suggests that even young infants display some form of metacognitive monitoring (Goupil and Kouider 2016; Goupil, Romand-Monnier, and Kouider 2016). Nevertheless, metacognitive abilities improve considerably across childhood, with the absolute accuracy of young children's metacognitive judgements and the impact of such judgements on information seeking being astonishingly low. For instance, in recent work, we showed that children's metacognitive judgements of their subjective uncertainty predicted information-seeking behaviour only in older but not in younger children (de Eccher, Mundry, and Mani 2024). What would be the role of children's increasing metacognitive awareness of their knowledge gaps in the current model? Furthermore, what would be the relationship between such advanced metacognitive awareness and the global error in the model proposed? Would increasing metacognitive awareness of areas of subjective uncertainty require changes to the model structure proposed in Figure 1? For instance, one possible outcome could be that increased metacognitive awareness of difficulties, for example, in processing text, leads to children with dyslexia disengaging with text at the earliest stages of processing highlighted above. In other words, with growing explicit awareness of their difficulties in text processing, children with dyslexia may inhibit even production of a reconstruction of presented textual input. Metacognitive awareness may in this case act like a firewall, inhibiting not just retrospective inattention but even early processing of imprecise sources, whilst more reliable sources filter through to early processing. This could then, potentially, suggest different pathways for rational attention and inattention whilst also speaking to previously reported differences in the proposed causes of some neurodivergent profiles, for example, deficits in early auditory processing or later working memory. This is but one of many solutions to integrating metacognition with information seeking. Clearly, this compelling theory of rational information seeking offers many avenues for bringing together the thinking from diverse fields, in terms of reconciling differences in neurodivergent and neurotypical information seeking but also across development and, potentially, the causes of certain neurodivergent profiles.

The author declares no conflicts of interest.