Shared neural geometries for bilingual semantic representations in human hippocampal neurons

This study reveals that the human hippocampus encodes language-independent semantic meaning not through shared individual neurons for equivalent words, but by preserving a common geometric organization of neural responses across languages via distinct readout axes.

The Gist

Imagine your brain is a massive, bustling library. Inside this library, there are millions of tiny librarians (neurons) whose job is to organize books (words) based on their meaning.

This study asked a fascinating question: If you speak two languages, like English and Spanish, do you have two separate libraries? Or is it one giant library where the books are just organized in two different ways?

The researchers went deep inside the human brain (specifically the hippocampus, the area responsible for memory and meaning) to watch these "librarians" in action while bilingual people listened to stories and spoke in both English and Spanish.

Here is what they found, explained through some simple analogies:

1. The "Rare Twins" (Cross-Language Neurons)

The researchers first looked to see if there were specific librarians who acted as "translators." They wondered: Is there a specific neuron that lights up for the word "dog" in English and the exact same neuron lights up for "perro" in Spanish?

They found a few of these "translator neurons." However, they were rare. Most neurons didn't care about the specific word; they cared about the context. So, the brain doesn't rely on a few special translators to switch languages.

2. The "Different Maps, Same Territory" (The Big Discovery)

This is the most important finding. The researchers discovered that while the individual librarians had different jobs in English versus Spanish, the overall layout of the library was identical.

The Analogy:
Imagine you have a map of a city.

• Map A (English): You draw the map with North at the top. The park is to the left of the library.
• Map B (Spanish): You draw the same city, but you rotate the map 90 degrees. Now North is on the right. The park is still to the left of the library, but the directions you use to get there are different.

The brain does this with meaning.

• In English, the brain uses a specific set of "directions" (neural axes) to locate the concept of "love."
• In Spanish, it uses a different set of directions to locate "amor."
• But, the distance between "love" and "hate" is the same in both maps. The distance between "cat" and "dog" is the same in both maps.

The brain preserves the geometry (the shape and distances between ideas) perfectly, even though the "compass" (the specific neurons firing) is rotated differently for each language.

3. The "Orchestra" Metaphor

Think of the brain's neurons as an orchestra.

• When you speak English, the orchestra plays a specific melody.
• When you speak Spanish, the same orchestra plays a different melody.
• The individual musicians (neurons) aren't playing the exact same notes in both songs. In fact, their tuning is quite different.
• However, the structure of the music is the same. The harmony, the rhythm, and the relationship between the notes remain consistent.

The brain uses the same group of musicians to play both languages, but it asks them to play from a different sheet of music (a different "readout axis"). This allows the brain to keep the two languages separate (preventing you from accidentally saying "perro" when you mean "dog") while still understanding that they mean the same thing.

4. Why This Matters: The "Zero-Shot" Magic

Because the brain keeps this perfect geometric shape, it can do something amazing: Predict meaning without knowing the translation.

If you know that in English, "wolf" is close to "dog" and far from "fork," your brain knows that in Spanish, "lobo" must be close to "perro" and far from "tenedor," even if you've never heard those Spanish words before. The brain uses the shape of the English map to guess the shape of the Spanish map.

The Takeaway

Your brain doesn't store English and Spanish in two separate filing cabinets. Instead, it builds one giant, high-dimensional "concept space."

• English is like looking at this space through a red filter.
• Spanish is like looking at it through a blue filter.

The objects (meanings) are the same, and their relationships to each other are identical. The only difference is the angle from which you are viewing them. This clever trick allows bilinguals to switch languages instantly without getting confused, because the "map" of meaning never changes—only the compass does.

Technical Summary

1. Problem Statement

Humans possess the remarkable ability to comprehend and express identical concepts in multiple languages without confusion. While neuroimaging (fMRI) suggests that bilinguals use overlapping neural networks for different languages, it remains unclear how the brain achieves two seemingly contradictory goals simultaneously:

1. Shared Meaning: Maintaining a unified representation of concepts across languages (e.g., "cat" and "gato" map to the same concept).
2. Language Separation: Preventing cross-language interference by keeping the languages functionally distinct.

The authors hypothesize that the brain does not rely on a simple "translation axis" or a set of dedicated "translation neurons." Instead, they propose that bilinguals utilize a shared high-dimensional semantic geometry (where the relative distances between concepts are preserved across languages) accessed via language-specific readout axes (rotations of the neural population activity). The hippocampus, known for its role in linking concepts, is the focal point of this investigation.

2. Methodology

Participants and Recording:

• Subjects: Four balanced bilingual patients (English/Spanish, acquired in early childhood) with treatment-resistant epilepsy.
• Recording Sites: Hippocampus.
• Techniques:
  • Patient 1-3: Stereo-electroencephalography (sEEG) with microelectrodes (AdTech Behnke-Fried).
  • Patient 4: Neuropixels 1.0 probe inserted into the exposed hippocampus during anterior temporal lobectomy.
• Total Data: 603 single neurons recorded during passive listening; 74 neurons during speech production tasks.

Experimental Paradigms:

1. Study 1 (Passive Listening): Patients listened to matched stories in English and Spanish (audiobooks and podcasts).
2. Study 2 (Speech Production): Patients read aloud 99 matched English-Spanish phrases.
3. Study 3 (Naturalistic Conversation): Unconstrained conversations with native speakers in both languages.

Data Processing & Analysis:

• Stimulus Alignment: Manual alignment of English and Spanish word pairs based on semantic equivalence (excluding non-equivalent grammatical structures).
• Semantic Embeddings: Contextual word embeddings were extracted using multilingual BERT (mBERT), specifically Layer 11 (identified as the layer with highest cross-lingual semantic alignment).
• Tuning Curves: Ridge regression was used to map shared semantic features (via joint PCA of mBERT embeddings) to neuronal firing rates for each language separately.
• Representational Geometry:
  • Representational Dissimilarity Matrices (RDMs): Computed using cosine distances between population firing rate vectors to capture the geometric structure of semantic space.
  • Subspace Alignment: Eigendecomposition of RSMs (1 - RDM) to identify principal semantic axes and calculate alignment indices between languages.
  • Local Prediction: A local Procrustes alignment method was used to predict a word's neural response in one language based on its neighbors in the other language's neural space.

3. Key Contributions & Results

A. Scarcity of "Cross-Language Neurons"

• The authors identified a small subset of neurons (approx. 5–19% depending on the task) that showed correlated firing rates for translation-equivalent words (e.g., "earth" and "tierra").
• Finding: While these "cross-language neurons" exist, they are rare. Crucially, removing them from the analysis did not disrupt the shared semantic geometry, indicating they are not the primary mechanism for cross-language translation.

B. Divergent Single-Neuron Tuning

• Finding: Individual neurons generally have distinct semantic tuning curves for English and Spanish.
• Evidence: The correlation between a neuron's English tuning vector and its Spanish tuning vector was significantly lower than the within-language reliability (split-half correlation).
• Implication: The brain does not use the exact same "recipe" (weights) to decode meaning in both languages at the single-neuron level.

C. Preservation of Population Geometry (The Core Finding)

• Finding: Despite divergent single-neuron tuning, the population-level semantic geometry is preserved across languages.
• Evidence:
  • Brain-Model Alignment: Neural RDMs in the hippocampus significantly correlated with mBERT semantic RDMs in both English and Spanish.
  • Cross-Language Alignment: The pairwise distances between words in English neural space strongly predicted the pairwise distances in Spanish neural space (significant positive correlations in RDMs).
  • Prediction: Using local geometric neighbors, the neural response of a word in Spanish could be accurately predicted from its English neural response (median angular error ~30–60°), significantly better than chance.

D. Language-Specific Readout Axes

• Finding: The shared geometry is accessed via different readout axes (rotations) for each language.
• Evidence:
  • Subspace alignment analysis showed that while the geometry (distances) is shared, the principal semantic axes (eigenvectors) of English and Spanish are less aligned than within-language subspaces.
  • However, the participation strength of individual neurons in the dominant coordination modes was highly correlated between languages.
• Conclusion: The same population of neurons is recruited for both languages, but the brain applies different linear readout weights (rotations) to extract language-specific meanings from the shared manifold.

4. Significance and Implications

• Mechanism of Translation: The study provides a neural mechanism for bilingualism that solves the "interference problem." By maintaining a shared semantic manifold but rotating the readout axes, the brain can access the same concepts without confusing the languages.
• Population Coding: It demonstrates that semantic representation in the hippocampus is an emergent property of population geometry rather than the sum of individual "concept cells."
• AI and Neuroscience: The findings align with how multilingual Large Language Models (like mBERT) operate, suggesting that high-dimensional geometric alignment is a fundamental principle of semantic representation in both biological and artificial intelligence.
• Hippocampal Role: It reinforces the hippocampus's role not just in episodic memory, but as a critical hub for flexible, language-independent conceptual modeling.

Conclusion: The bilingual brain encodes a language-independent internal model for meaning within the hippocampus. This model is implemented as a shared high-dimensional geometric structure where concepts maintain consistent relative distances across languages, accessed through distinct, language-specific neural readout configurations.