Metabolic constraints restricting theory development

Unifying theories across Marr’s levels of analysis is a laudable goal (Griffiths, Reference Griffiths2026, p. 294), but it is not always immediately beneficial. At some point, metabolic realities must be faced, but putting such concerns into the “too hard basket” is often practical in the short term. The history of backpropagation in artificial neural networks illustrates this point.

Backpropagation – the algorithm used to train multilayered neural nets – is the bedrock of deep learning. Real neurons, however, lack the symmetric weights and global error signals that standard backpropagation requires (see Whittington & Bogacz, Reference Whittington and Bogacz2019), leading some to conclude that such models bear no relevance to the brain (Crick, Reference Crick1989; Grossberg, Reference Grossberg1987). Rather than abandoning such approaches to understanding the mind, however, many cognitive scientists continued using backpropagation-based models to develop models of human intelligence, either placing implementation concerns on the backburner or working in parallel on biologically feasible yet functionally comparable models (e.g., Hinton & McClelland, Reference Hinton, McClelland and Anderson1988; O’Reilly, Reference O’Reilly1996). Recently, several biologically plausible alternatives have been proposed that mimic the functional benefits of backpropagation (for reviews, see Lillicrap et al., Reference Lillicrap, Santoro, Marris, Akerman and Hinton2020; Whittington & Bogacz, Reference Whittington and Bogacz2019).

In a sense, this brief history vindicates Haueis & Colaço’s proposal in that theory was improved by adapting models to known biological constraints. However, if the field had entirely dismissed “connectionist” models because of their biological implausibility, we may have forgone the foundation upon which better theories were eventually built. Scepticism (much of it warranted) toward neural networks as models of the mind persists today, driven by the enormous differences in energy consumption between the brain and AI models (section 4.3). It is arguable, however, that the impressive capabilities of large-scale AI – which demonstrate that complex, human-like reasoning is achievable in artificial systems – while not intended to serve as models of the mind, have spurred interest in developing biologically constrained alternatives. Metabolic implausibility tells us that a model must be improved upon, but it should not lead to the abandonment of a particular approach at an early stage, especially when existing alternatives do a poorer job of explaining the data (e.g., category learning in symbolic vs. connectionist models).

The weight placed on whether a model aligns with the brain’s biology should be considered with respect to the model’s maturity. If a model is widely accepted, with differences between formalised mechanisms being relatively minor (e.g., differences between diffusion decision models; Ratcliff et al., Reference Ratcliff, Smith, Brown and McKoon2016), then metabolic concerns will be useful in helping adjudicate between accounts. If, however, the theories being debated predominantly work at Marr’s computational level (e.g., opportunity cost vs. resource-based accounts of fatigue), or the behaviour theorised is so complex that we have little idea how the brain may implement the hypothesised algorithms (e.g., models of social cognition, Alister & Perfors, Reference Alister and Perfors2025; Young & de-Wit, Reference Young and de-Wit2025), then biological concerns are best deferred until our theories are sufficiently advanced.