A better pair of looking glasses

Most researchers, whatever side they come from, will agree functional specialization is a matter of degree not kind. Now instead of arguing about whether syntactic processing is localized or distributed, a point of contention is more of which approach is more productive. Or in other words, what is debatable is less about the view outside the window, but more about the glasses we used to look at the view. Imagine that we observe the same landscape through different types of looking glasses—one with a narrow, zoomed-in lens and the other with a wide-angle, zoomed-out lens. Both are looking at the same scene (the brain’s processing of syntactic information), but the narrow lens provides a close-up view of individual trees (localized brain regions), while the wide-angle lens shows the entire forest at once (distributed brain networks).

When operationalized, the distributed approach of syntax in discussion is often characterized by firstly, a spatial resolution at region or node level; secondly, a narrative of how one single syntactic operation is implemented by many regions differentially (e.g. subject or object extraction, wh-movement, integration of word lists into sentences). Admittedly, this body of work has produced systematic and solid evidence (Blank et al. 2016; Christensen 2008; Shain et al. 2024). However, I argued that, the pure amount of positive evidence, may not be sufficient enough to justify the preference of one approach over another. Inductive reasoning can enter a self-reinforcing mode by deriving conclusions from known cases and it’s easy to find consistent evidence through stereotypical practices which only allow for certain interpretations of the not-so-objectively-neutral evidence (Popper 2002). Then, what should be the proper metric of comparison here?

From the perspective of scientific instrumentalism, theories, as tools for organizing observations, should be evaluated on the basis of its adequecy for making accurate, testable predictions about future observations (Fraassen 1980). If a theory consistently predicts phenomena correctly, it’s considered successful—even if we’re unsure whether it describes reality. That is to say, the modular approach may risk oversimplifying the brain’s functioning, but as long as it allows for better predictions, it should be preferred over the other approach.

The same line of thinking is also reflected in Lakatosian research programmes where the value of scientific approaches is determined by whether they are progressive or degenerative (Doerig et al. 2023). Progressive research programmes generate new insights while degenerative ones descend into a repeated corroboration of existing ideas. In the context of syntactic processing, the logic is similar. Then, the question becomes, which approach is progressive, i.e. more productive in generating new, testable prdictions?

Both approaches, the modular approach which sounds like “region A the main region for doing X” and the distributed approach that sounds like “region A, B, C are all doing X in varying degrees,” are still coarse-grainded when we are aligning them with the high-resolution descriptions of syntax by linguists. But the mismatch in granularity is at least less pronounced, as the functions of discrete modules are more clearly defined. The implication is that, if our ultimate goal is to generate predictions in terms of the computations carried out by the brain, modular findings will make it easier to build up complexity incrementally to align with linguists’ specification of syntactic operations. In fact, computational models that start with modular assumptions but gradually incorporate distributed dynamics have demonstrated notable successes. For example, in modeling syntactic structure recognition, one study adopted a discrete, three-layer architecture that approximates the structure of cerebellum, with each layer corresponding to cells with different biological functions and has successfully captured the emergence of syntactic recognition in the middle layer (Ohmae and Ohmae 2024). Another computational model of reading, though framed within a connectionist architecture, similarly builds upon functionally dissociated visual, orthographic, phonological modules to account for dyslexia (Chang et al. 2024).

However, I am not suggesting that, distributed assumptions will takes us nowhere. What I intend to highlight is that starting with distributed assumptions of syntactic processing without a modular foundation can be ineffective in generating predictions. The vague similarity between it and artificial neural networks is still very far from explaining the computations of the brain.

Reference:

Blank, I., Balewski, Z., Mahowald, K., & Fedorenko, E. (2016). Syntactic processing is distributed across the language system. NeuroImage, 127, 307–323. https://doi.org/10.1016/j.neuroimage.2015.11.069

Chang, Y.-N., Welbourne, S., Furber, S., & Lambon Ralph, M. A. (2024). Simultaneous simulations of pure, surface and phonological acquired dyslexia within a full computational model of the primary systems hypothesis. Cortex; a Journal Devoted to the Study of the Nervous System and Behavior, 179, 112–125. https://doi.org/10.1016/j.cortex.2024.07.006

Christensen, K. R. (2008). Interfaces, syntactic movement, and neural activation: A new perspective on the implementation of language in the brain. Journal of Neurolinguistics, 21(2), 73–103. https://doi.org/10.1016/j.jneuroling.2007.01.002

Doerig, A., Sommers, R. P., Seeliger, K., Richards, B., Ismael, J., Lindsay, G. W., Kording, K. P., Konkle, T., van Gerven, M. A. J., Kriegeskorte, N., & Kietzmann, T. C. (2023). The neuroconnectionist research programme. Nature Reviews Neuroscience, 24(7), 431–450. https://doi.org/10.1038/s41583-023-00705-w

Fraassen, B. C. van. (1980). Arguments Concerning Scientific Realism. In Bas. C. van Fraassen (Ed.), The Scientific Image (p. 0). Oxford University Press. https://doi.org/10.1093/0198244274.003.0002

Ohmae, K., & Ohmae, S. (2024). Emergence of syntax and word prediction in an artificial neural circuit of the cerebellum. Nature Communications, 15(1), 927. https://doi.org/10.1038/s41467-024-44801-6

Popper, K. (2002). The Logic of Scientific Discovery (2nd edition). Routledge. https://doi.org/10.4324/9780203994627

Shain, C., Kean, H., Casto, C., Lipkin, B., Affourtit, J., Siegelman, M., Mollica, F., & Fedorenko, E. (2024). Distributed Sensitivity to Syntax and Semantics throughout the Language Network. Journal of Cognitive Neuroscience, 36(7), 1427–1471. https://doi.org/10.1162/jocn_a_02164