The high-level hypothesis

More anatomically specialised architectures are defined as those whose physical structures map to a narrower range of system-relevant computational tasks. For instance, the autoassociation capability of the hippocampus is more general than retinal cell edge detection but less computationally general than predictive processing in the neocortex. Detailed analysis can leverage methods to quantify dimensions of “computational generality,” e.g., programmability (Zenil, Reference Zenil2013), VC dimension (Vapnik, Reference Vapnik1998), and Rademacher complexity (Bartlett & Mendelson, Reference Bartlett and Mendelson2002).

“Metabolic efficiency” can be operationalised by adapting “information-energy efficiency”: average ATP count to process environmentally representative information in one functional input-output sweep on a system-relevant task (e.g., 1x image recognition, 1x learning update).

When two systems have been available in evolutionary terms and are both capable of a given computation, the hypothesis claims the brain will favour the more specialised system due to its greater efficiency. In practice, the brain and its evolutionary environment are highly complex, path-dependent systems (Jacob, Reference Jacob1977), such that other factors might dominate. In general, the hypothesis should be stronger under greater selection pressure – phylogenetically older systems, more frequent tasks, etc. – pointing to possible cases to explore first.

Specific cases in inter-neuronal networks to explore

The brain has evolutionary access to a highly general-purpose architecture and yet exploits it rarely. Fully connected single hidden-layer neural networks can approximate any function (Hornik, Stinchcombe, & White, Reference Hornik, Stinchcombe and White1989). Networks with identical neurones and uniform connectivity structure represent very low levels of specialisation. However, even 20%-30% connectivity qualifies as densely connected local circuits in the brain (Perin, Berger, & Markram, Reference Perin, Berger and Markram2011; Song et al., Reference Song, Sjöström, Reigl, Nelson and Chklovskii2005), and fully connected cliques occur only at very small scales (Reimann et al., Reference Reimann, Nolte, Scolamiero, Turner, Perin, Chindemi and Markram2017). Instead, the most computationally general brain structure, the neocortex, exploits layers and reciprocal dynamics in highly structured networks (Mountcastle, Reference Mountcastle1997). The hypothesis suggests that metabolic efficiency analysis would locate this level of structural fine-tuning in an approximate Goldilocks zone, given the computational generality of target tasks.

The hierarchy of anatomical specialism extends beyond structured connections between neurones to specialised neuronal cells. The cerebellum implements narrow adaptive filtering tasks with efficient specialised microcircuitry via Purkinje neurones and climbing fibre error signals (Albus, Reference Albus1971; Kawato et al., Reference Kawato, Ohmae, Hoang and Sanger2021; Marr, Reference Marr1969). Standard predictive coding in cortical microcolumns (Bastos et al., Reference Bastos, Usrey, Adams, Mangun, Fries and Friston2012) could implement these tasks in principle, but this hypothesis suggests the cerebellum is favoured (in part) for efficiency reasons, e.g., multiple microcolumns might be required for each Purkinje cell.

Cases beyond inter-neuronal synaptic connectivity?

Neuronal signalling is metabolically expensive, so offloading computational primitives to non-neuronal cells (where viable) is likely to be even more favourable under this hypothesis than neuronal specialism.

Computationally narrow tasks in phylogenetically older systems provide candidate cases to examine. For instance, Müller glia in the retina play roles in signal processing, guiding light to photoreceptors and regulating neurotransmitter dynamics (Newman & Reichenbach, Reference Newman and Reichenbach1996; Reichenbach & Bringmann, Reference Reichenbach and Bringmann2013), while Bergmann glia in the cerebellum interact with synaptic timing and plasticity (Bellamy, Reference Bellamy2006; De Zeeuw & Hoogland, Reference De Zeeuw and Hoogland2015).

A metabolic comparison of neuronal and non-neuronal mechanisms for implementing these computational primitives should favour the latter. In cases where a non-neuronal efficiency gain has not been exploited, we expect a countervailing reason: perhaps the efficiency gain is too modest relative to robustness priorities (Goldman et al., Reference Goldman, Golowasch, Marder and Abbott2001), the task too infrequent for significant selection pressure, the required system transition too large to have occurred (yet) by chance mutation, or system reuse more efficient than specialisation overall (e.g., once construction and system interaction costs are considered; see Anderson, Reference Anderson2010).

This hypothesis also motivates attention to emerging, more speculative avenues of research on candidate analogue computational mechanisms, inspired also by the parallels to computing R&D in Haueis and Colaço.

Consider the emerging evidence on ephaptic coupling mechanisms in mammal brains, given how modest field strengths shape activation phases (e.g., 0.2 to 2 mV/mm) (Anastassiou et al., Reference Anastassiou, Perin, Markram and Koch2011; Anastassiou & Koch, Reference Anastassiou and Koch2015; Jefferys, Reference Jefferys1995) and extracellular fields generated by neurones shape both activations (Pinotsis & Miller, Reference Pinotsis and Miller2023) and working memory (Pinotsis & Miller, Reference Pinotsis and Miller2022). Since EM fields are already produced by neurones, it is potentially highly efficient to leverage them computationally. Industrial R&D also leverages fields as energy-efficient analogue mechanisms in their own right – e.g., matrix multiplication (Bai et al., Reference Bai, Xu, Tan, Sun, Li, Wu, Morandotti, Mitchell, Xu and Moss2023), Fourier transforms (Heinz, Artman, & Lee, Reference Heinz, Artman and Lee1970), and combinatorial optimisation (e.g., Hamerly et al., Reference Hamerly, Inagaki, McMahon, Venturelli, Marandi, Onodera, Ng, Langrock, Inaba, Honjo, Enbutsu, Umeki, Kasahara, Utsunomiya, Kako, Kawarabayashi, Byer, Fejer, Mabuchi and Yamamoto2019; US patent US11698945B2) – providing potential (if highly speculative) inspiration for applications to investigate in the brain.

Components have lower metabolic costs than their host systems, motivating an examination of intra-neuronal mechanisms. A classic example is dendritic computation enhancing neurone-level capabilities (London & Häusser, Reference London and Häusser2005; Ujfalussy et al., Reference Ujfalussy, Makara, Branco and Lengyel2015). Turning to even smaller subsystems, a speculative example is leveraging extremely rapid path integral mechanics for predictive processing in cases where Hodgkin–Huxley neurones are too slow (Wiest & Puniani, Reference Wiest and Puniani2025). Much more work is needed to assess any role that quantum phenomena might play in cognition, but the R&D parallel provides further motivation to investigate the possibility, given the energy efficiency argument for quantum computers (Meier & Yamasaki, Reference Meier and Yamasaki2025; Nielsen & Chuang, Reference Nielsen and Chuang2010; Preskill, Reference Preskill2018; US patent US11238000B2).

As appreciation of evolutionary complexity improves, it is plausible that efficiency gains motivating industrial R&D might have also been mined by evolution. Metabolic considerations provide a structured framework for investigating these possibilities and for converting intriguing speculations into disciplined science.