Cognition as a Protean Endeavor

Cognitive neuroscience, shifting heuristics, and the myth of converging truths.

Engraving of Proteus from Pierre Mussard’s Historia deorum fatidicorum (1675)

In Greek mythology, Proteus is a sea god who knows past, present, and future, truths but refuses to share them. To get an answer out of him, you have to ambush him and hold him down. The moment you grab him, he tries to slip away by shifting shapes, turning for example into a dragon, a bird, a serpent, or a tree. If you hold on tightly enough through all of those transformations, Proteus eventually tires out, returns to his true form, and speaks the truth.

All the way in the seventeenth century, Francis Bacon used Proteus as a metaphor for science. Arguing that nature truly is shifting and elusive. Bound by the constraints of an experiment, however, nature is forced to stop shape shifting and reveal its single underlying form. But who is to say such a thing exists?

Much of modern cognitive neuroscience still runs on this assumption, called the myth of converging truths. On this assumption we track blood flow with fMRI, record electrical rhythms with EEG, and measure reaction times in behavioral tasks. Each method gives a partial picture, that will lead us further in the search for emergent patterns of cognition. If we just had better tools and larger datasets, our findings will converge onto one canonical blueprint of the human mind.

But what if the brain does not play by the second half of the Proteus myth? What if, when we hold cognition down in a laboratory task, the shape it takes is not its final resting form at all, but just a temporary shape forced by how we are holding it? There is no saying that we are not being tricked by proteus all the same, not because we never managed to force him into his true form, but simply because he can’t. What if proteus has no true form?

Studying the River Without Freezing the Water

Human cognition could be inexhaustibly complex, if you ever try to pin it down on a representation, that is always just that, a representation of the reality, and there is no saying whether these representations ever converge to one unary form of cognition. That is to say, not only might mechanisms of cognition between participants differ vastly, perhaps the brain is made from an ever-changing, heuristics-driven mechanism that has no singular form of representation.

Once you take that possibility seriously, many stubborn problems in neuroscience start to look less like measurement noise and more like core features of how biological minds work.

Consider how we compare brains across people. Standard neuroimaging studies warp dozens of individual brain scans into a shared template and average their activity together. Sure, any respecting practicioner is aware of inter-subject variance. We know averaging their brain activity does not reveal the one true representation of working memory, yet we still keep on trying. Because we like our heuristic, we like our simple models of the world (why do you think we invented religion). Research has well established that group-averaging blurs out distinct, reliable network layouts that are unique to each person. But we are not occupied with fundamental truth. Yet much so the same as is the case for religion, once we chose a story (albeit a usefull one) we risk closing ourselves off to different narratives.

Biology rarely insists on a single wiring diagram. In systems biology, Edelman and Gally (2001) described this as degeneracy: structurally different neural circuits can produce the exact same behavioral output. Prinz, Bucher, and Marder (2004) demonstrated this in the crustacean stomatogastric circuit (a 30-neuron digestive circuit that controls gut rhythms in crabs), showing across millions of network configurations that indistinguishable rhythmic activity can arise from vastly different combinations of synaptic strengths and ion channels.

If that is true for a tiny circuit in a crab, scale it up to eighty-six billion neurons in humans shaped by individual lives and habits.

Even within a single brain over time, representations refuse to sit still.

We often talk about neural representations as if they were static files on a hard drive (One numerous amounts of the technical methaphors of the human mind we insist on using). Yet longitudinal recordings have uncovered a phenomenon known as representational drift. When mice perform a well-learned virtual navigation task over weeks (Driscoll et al., 2017), or smell the exact same odor across days (Schoonover et al., 2021), their behavior stays stable while the specific neurons encoding that place or odor in parietal or olfactory cortex continuously shift.

Get Noah Wijnheijmer’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

Why would a brain drift like this instead of locking in one permanent code? One plausible reason is that the brain is not trying to maintain a single, stationary mirror of reality. As Gigerenzer and Gaissmaier (2011) argue, cognition often works less like a general-purpose computer and more like an adaptive toolbox of heuristics: fast, context-sensitive rules of thumb pieced together for the immediate environment. What we capture in a scanner and call a representation may simply be the footprint of a temporary heuristic the brain improvised for that task.

What Does This Mean for Computational Neuroscience?

In recent years, researchers have shown that the internal activations of deep vision and language models can predict human brain activity remarkably well during shared tasks (Schrimpf et al., 2021). In machine learning, Huh and colleagues (2024) framed a related trend as the Platonic Representation Hypothesis: as models scale up across different modalities, their internal representations appear to converge toward a shared statistical structure because they are all modeling the same physical world.

There is real merit to that argument. Human brains do not operate in a vacuum; we share evolutionary constraints, navigate the same physical world, and share the same thinking template we call language. Some shared structure has to exist for communication or cross-subject decoding to work at all.

Still, the protean view reminds us to be careful about what model-to-brain alignment actually proves. When two flexible systems are funneled through the same narrow task, a linear regression can often find shared variance between them, even if their underlying mechanisms differ.

I am not argueing that viewing cognition as protean makes traditional computational modeling pointless, merely what questions we should be asking ourselves. Instead of only asking which single architecture best matches the average brain, we can model how a brain switches heuristics when contexts shift. If you acknowledge that cognition is flowing, you can be prepared to go with that flow, and adapt.

Proteus is a useful reminder. In cognitive and computational neuroscience, we have to hold the mind down with tasks and models just to take a measurement. The key is remembering that the shape we catch in our hands is only one of the things cognition can be, and that the real marvel may not be the form it takes when held still, but how effortlessly it keeps changing.

References

1. Pesic, P. (1999). Wrestling with Proteus: Francis Bacon and the “Torture” of Nature. Isis, 90(1), 81 to 94. https://doi.org/10.1086/384242
2. Gordon, E. M., Laumann, T. O., Gilmore, A. W., et al. (2017). Precision Functional Mapping of Individual Human Brains. Neuron, 95(4), 791 to 807. https://doi.org/10.1016/j.neuron.2017.07.011
3. Edelman, G. M., & Gally, J. A. (2001). Degeneracy and complexity in biological systems. Proceedings of the National Academy of Sciences, 98(24), 13763 to 13768. https://doi.org/10.1073/pnas.231499798
4. Prinz, A. A., Bucher, D., & Marder, E. (2004). Similar network activity from disparate circuit parameters. Nature Neuroscience, 7(12), 1345 to 1352. https://doi.org/10.1038/nn1352
5. Driscoll, L. N., Pettit, N. L., Minderer, M., Chettih, S. N., & Harvey, C. D. (2017). Dynamic Reorganization of Neuronal Activity Patterns in Parietal Cortex. Cell, 170(5), 986 to 999. https://doi.org/10.1016/j.cell.2017.07.021
6. Schoonover, C. E., Ohashi, S. N., Axel, R., & Fink, A. J. P. (2021). Representational drift in primary olfactory cortex. Nature, 594, 541 to 546. https://doi.org/10.1038/s41586-021-03628-7
7. Gigerenzer, G., & Gaissmaier, W. (2011). Heuristic Decision Making. Annual Review of Psychology, 62, 451 to 482. https://doi.org/10.1146/annurev-psych-120709-145346
8. Schrimpf, M., Blank, I. A., Tuckute, G., et al. (2021). The neural architecture of language: Integrative modeling converges on predictive processing. Proceedings of the National Academy of Sciences, 118(45), e2105646118. https://doi.org/10.1073/pnas.2105646118
9. Huh, M., Cheung, B., Wang, T., & Isola, P. (2024). The Platonic Representation Hypothesis. Proceedings of the 41st International Conference on Machine Learning (ICML). https://arxiv.org/abs/2405.07987
10. Kriegeskorte, N., & Diedrichsen, J. (2019). Peeling the Onion of Brain Representations. Annual Review of Neuroscience, 42, 407 to 432. https://doi.org/10.1146/annurev-neuro-080317-061906