Distributed Semantic Mapping Across the Human Cortex

89 concepts64 papers1 proposal

For decades, the dominant model of language in the brain was modular. We believed that specific "boxes"—like Broca’s area for production and Wernicke’s area for comprehension—handled isolated tasks. However, recent research using naturalistic stimuli and high-dimensional modeling suggests a more fluid reality: linguistic meaning is not localized, but is instead represented as a distributed semantic map spanning nearly the entire cortex.

The Shift to Naturalism

The intuition behind this shift comes from how we study the brain. Traditional experiments used "toy" stimuli—isolated words or simple sentences—which often failed to trigger the brain's full processing capacity. When Leila Wehbe et al. moved toward naturalistic paradigms, such as having subjects read a chapter of Harry Potter in an fMRI scanner, they discovered that language processing is far more bilaterally distributed than previously thought.

This distributed nature is "amodal." As Fatma Deniz et al. demonstrated in The Representation of Semantic Information Across Human Cerebral Cortex During Listening Versus Reading Is Invariant to Stimulus Modality, the brain maps the meaning of a story in essentially the same way whether you are reading it or listening to it. The semantic map is a stable, large-scale functional architecture that integrates information across the temporal, parietal, and prefrontal cortices.

The Entanglement of Syntax and Semantics

One of the most striking findings in this research direction is that the brain does not appear to have a dedicated "grammar module." Instead, syntactic structure and semantic meaning are deeply entangled.

Aniketh Janardhan Reddy et al. showed in Can fMRI reveal the representation of syntactic structure in the brain? that there is substantial spatial overlap between regions processing structure and those processing meaning. To prove this, Shaonan Wang et al. used a "feature elimination" method to surgically remove syntactic information from AI embeddings. They found that the ability to predict brain activity dropped across a broad, distributed network, suggesting that the biological implementation of grammar relies on the same overlapping substrates as meaning.

Bilateral Recruitment and Complexity

While lexical processing (individual words) shows some left-hemisphere bias, the transition to complex, multi-word composition triggers a massive expansion in cortical recruitment. Alexandre Gramfort et al. found that compositional syntax and semantics are processed bilaterally. The right hemisphere, often dismissed in older models, appears essential for handling the "heavy lifting" of language:

  • Coarse Coding: Yang Peng argues the right hemisphere specializes in "coarse coding" for distant or novel associations, such as metaphors.
  • Narrative Tracking: Leila Wehbe et al. identified significant right-hemisphere activity for tracking plot and high-level narrative features.

AI as a Mirror for Distributed Representations

The reason modern Large Language Models (LLMs) like BERT align so well with human brain activity is that they share this distributed philosophy. In Intriguing properties of neural networks, the authors observed that semantic information in deep networks is spread across the entire activation space rather than being localized in individual "concept neurons."

This mirrors the brain's "holographic" organization. Shinji Nishimoto et al. describe this as continuous semantic mapping, where information is represented as an organized distribution across the cortex. By treating the brain as a high-dimensional vector space—much like an LLM—researchers can now decode complex thoughts. For instance, Qingyao Ai et al. developed BrainLLM, which treats fMRI signals as a "modality" to reconstruct the language a person is hearing, proving that the distributed signals in the prefrontal and auditory cortices contain enough information to drive generative language models.

Why It Works: The Power of the Whole

The core intuition of this field is that distribution provides robustness. Models that focus only on "hotspots" or localized regions often lose critical information. As S. Nishida et al. argued, aggregating activity across the entire cortical surface better reflects human semantic perception than focusing on the most predictable voxels. By spreading information across overlapping networks, the brain (and the models we build to mimic it) can represent an almost infinite variety of concepts and relationships within a finite physical space.

Key Foundational Work

Resolving sentence ambiguity with planning and working memory resources: Evidence from fMRI (2007) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_distributed_semantic_representation_fcbf60c2,
  title = {distributed semantic representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/distributed-semantic-representation/rdr_fcbf60c23dce445da9901b11e265a046},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_distributed_semantic_representation_fcbf60c2_v1,
  title = {distributed semantic representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/distributed-semantic-representation/rdr_fcbf60c23dce445da9901b11e265a046/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake