‘Fix’ the Child or Change the Learning Environment?

Each of us experiences constraints upon how we process the world around us. In turn, we each allocate resources strategically, favouring channels of input that minimise the epistemic uncertainty about that world. According to this view emergent differences in cognition need not be attributed to nebulous information processing ‘deficits’, but to the natural unfolding of resource-rational behaviour over developmental time. In their formalisation of ‘Rational Inattention’, Jones, Jones, Koldewyn and Westerman, tie together many key developmental observations. Perhaps most crucially, they attempt to reconcile two seemingly disparate phenomena: exogenous sources of uncertainty can actively attract our attention (e.g., Poli et al. 2020), whereas, endogenously driven uncertainty stemming from noise within the system, drives disengagement from that source of input (e.g., Jones et al. 2018).

The rational allocation of resources, within the context of input channels with differing levels of noise, is implemented within simulations. Artificial agents disengage from sources of input that they do not gain information from. This is formalised as a loss function, summing the losses derived from a four-path autoencoder. Crucially, imprecision (or noise) within one of those input paths makes it difficult to compress and subsequently decode information. This creates an information bottleneck (Tishby, Pereira, and Bialek 2000). Put simply the agent does not update its learnable parameters on the tasks it cannot solve. Instead, the global signal is better minimised by referencing the inputs that have greater endogenous informational precision.

Rational Inattention is a theoretical framework centred on the dynamic interaction between the child and (potential) domains of information, and how the allocation of resources might shape future processing. We like many of its features. But in this short commentary we would like to focus on what this account does not yet explain. Even if we side-step the obvious issue of what makes information differentially judged to be precise or not in the first place, how well does the authors’ characterisation of neurodevelopmental differences match reality?

As the authors acknowledge, some radical simplification is necessary in order to get the simulations running, but there are meaningful differences between real neurodevelopmental phenotypes and the simulations. First, discrete domain-specific difficulties are pretty rare in reality. Moreover, where these difficulties do occur they commonly cascade to difficulties in other domains, rather than driving enhanced processing of other streams (e.g., Goh, Griffiths, and Norbury 2021). Second, while areas of strength and difficulty do occur (e.g., Astle et al. 2019), they tend to be relative areas of strength within an individual, rather than absolute strengths. In short, the biggest differences between individuals tend to be domain general, rather than restricted to specific domains. These two observations seem hard to reconcile with a model in which profiles emerge from specific processing bottlenecks within restricted, independently summed, input channels. Instead, it seems likely that the complex reality at least partially reflects the interactions between those processing channels. There are numerous demonstrations that initial subtle differences across several domains cohere into the classic diagnostic traits over several years, not usually becoming a clearly identifiable phenotype until later (e.g., Dawson, Rieder, and Johnson 2023). How might we extend the concept of rational inattention to capture the domain general, cascading and wide-spread subtle differences that we actually observe ‘in the wild’?

Related to these issues, how well do the practical implications of the framework extend to reality? Rational inattention situates the child as an active learner. There is a crucial role for the environment, or more particularly the child's interaction with that environment, in shaping the emergence of their cognitive system. We agree whole-heartedly; this is a necessary ingredient of any good developmental theory. However, one key practical implication that the authors deduce from their account is that formal programmes to facilitate deep, structured engagement with difficult information sources might yield clinical benefits. We would advance an alternative interpretation. The success of interventions based on the principle of forcing children to engage with difficult to process materials, however ‘deep or structured’, tends to be poor. As a field, we have tried to train particular cognitive skills, but the improvements are restricted to whatever is trained (e.g., Ganesan et al. 2024) and do not seem to drive better engagement with sources of input in the real world; we have trained autistic children to get really good at recognising emotion in faces, but with no evidence for knock-on changes in wider social skills (Zhang et al. 2021); we have trained working memory in kids with ADHD, with zero wider benefit (Cortese et al. 2015). It may be a controversial view, but interventions designed to hammer apparent ‘deficits’ yield minimal widespread benefit, and can themselves drive both stereotyping and stigma surrounding different diagnostic categories. An alternative approach is to consider how we can harness relative strengths. Our view is that rather than using this framework to motivate attempts to ‘remediate’ differences in cognitive profile, we should instead use it to think about how we structure learning such that different profiles do not become barriers to learning.

In summary, Rational Inattention as outlined by Jones et al. (this issue) is a real step toward a model that formalises how the emergence of cognitive profiles partly reflects of resource-rational behaviour. In our view, the next crucial steps are to consider whether or how we can take this concept forward, presumably alongside other core developmental tenets, to better explain the cognitive profiles we see in actual children, and why some types of support work better than others.

The authors declare no conflicts of interest.