A population code for semantics in human hippocampus

This study demonstrates that the human hippocampus encodes word meanings through distributed population activity patterns that align with contextual semantic embeddings, revealing distinct coding principles for word frequency and a multidimensional semantic subspace that supports robust meaning tracking during speech comprehension.

The Gist

Imagine your brain is a massive, bustling library. For a long time, scientists thought that when you heard a word like "apple," a single, specific librarian (a neuron) would jump up and shout, "I know that word!" This is called a "one-to-one" code.

But this new study suggests the brain works more like a symphony orchestra or a complex dance troupe. No single musician plays the whole song; instead, the meaning of the word comes from the specific pattern of how hundreds of musicians play together at the same time.

Here is a breakdown of what the researchers found, using simple analogies:

1. The "Mixed-Selectivity" Orchestra

The researchers listened to the brains of epilepsy patients (who had tiny microphones implanted in their brains to monitor seizures) while they listened to stories. They found that the hippocampus (a brain region famous for memory) doesn't have a "word for apple" neuron.

Instead, think of a single neuron like a chameleon.

• One neuron might react to "apple," but it also reacts to "running," "blue," and "yesterday."
• Another neuron might react to "apple," but also "justice," "loud," and "tomorrow."
• The Analogy: Imagine a word is a specific color. No single paintbrush holds that exact color. Instead, the color is created by mixing a specific recipe of red, blue, and yellow from many different brushes. To understand the word "apple," your brain looks at the unique "recipe" of activity across hundreds of neurons, not just one.

2. Context is King (The "Sharp" Knife vs. The "Sharp" Mind)

Language is tricky. The word "sharp" means something different in "a sharp knife" than in "a sharp mind."

• Old Idea: The brain might have a static file for "sharp" that never changes.
• New Finding: The brain updates the "recipe" every time. When you hear "sharp knife," the orchestra plays one specific song. When you hear "sharp mind," the same musicians play, but they change their volume and timing to create a completely different song.
• The Analogy: It's like a chameleon changing its skin. The animal is the same, but its appearance changes based on the background. The brain changes the neural pattern based on the story's context.

3. The "AI" Connection

The researchers compared the brain's activity to modern AI language models (like GPT-2 and BERT).

• Static AI (Word2Vec): Imagine a dictionary where "bank" always means the same thing, whether it's a river bank or a money bank. The brain doesn't work like this.
• Contextual AI (GPT-2/BERT): These AIs understand that "bank" changes meaning based on the sentence. The researchers found that the brain's "orchestra" patterns matched these smart, contextual AIs much better than the old dictionary-style AIs.
• The Takeaway: Your brain is a context-aware storyteller, not just a dictionary look-up.

4. The "Crowded Room" vs. The "Quiet Corner"

The study found something surprising about common words (like "the," "and," "you") versus rare words (like "squirrel," "umbrella").

• Rare Words: When you hear a rare word, the brain creates a very distinct, unique pattern. It's like a quiet corner in a library where only a few people are talking. It's easy to tell them apart.
• Common Words: Common words are used in so many different ways (high "polysemy") that their neural patterns are all over the place. They are like a crowded, noisy party. Because "the" can mean so many different things depending on the sentence, the brain's pattern for "the" is messy and overlaps with many other patterns.
• The Twist: To make sure you don't get confused by words that sound or mean almost the same thing (like "brother" and "sister"), the brain actually uses a contrastive code. It deliberately makes the patterns for very similar words more different from each other than you'd expect, just to be safe. It's like the brain saying, "These two words are neighbors, so let's put a big fence between them so we don't mix them up!"

5. The Big Picture

The study concludes that the brain doesn't store meaning in single cells. Instead, meaning is a distributed landscape.

• The Map: Imagine a map where every word is a city. Words with similar meanings are cities close together. Words with different meanings are far apart.
• The Journey: When you hear a word, your brain doesn't just visit one city; it lights up a whole region of the map. The "shape" of that light-up region tells the brain exactly what the word means in that specific moment.

In short: Your brain is a dynamic, context-sensitive orchestra. It doesn't use a single note to represent a word; it uses a complex, shifting chord that changes depending on the story, ensuring that even the most similar words can be told apart.

Technical Summary

1. Problem Statement

The fundamental principles by which the human brain encodes word meaning remain poorly understood. While previous research suggested that specific neurons might act as "grandmother cells" (tuned to single concepts) or that semantic categories are represented by distinct neural populations, these models struggle to explain the fluidity and context-dependency of natural language.

• The Gap: Large Language Models (LLMs) have demonstrated that semantic relationships are best captured by distributed population geometry (vector embeddings) rather than isolated features. However, it is unclear if the human brain, specifically the hippocampus (a region critical for memory and relational processing), utilizes a similar distributed, context-dependent population code.
• The Challenge: Assessing this hypothesis is difficult because language is highly contextual (e.g., "sharp" means different things in "sharp knife" vs. "sharp mind"), and most neuroscience studies have relied on static, non-contextual word representations or discrete stimuli rather than continuous narrative speech.

2. Methodology

The study employed intracranial electrophysiology combined with advanced natural language processing (NLP) techniques.

• Participants & Data Collection:
  • Subjects: 10 adult epilepsy patients undergoing intracranial monitoring (stereotactic EEG, sEEG).
  • Recording: Single-unit and multi-unit activity recorded from the hippocampus (356 neurons total) using microwire probes.
  • Stimuli: Participants listened to 47 minutes of continuous narrative speech (6 episodes from The Moth podcast, 7,346 words). A subset of 4 patients also listened to "Jabberwocky" stimuli (nonsense words preserving syntax but lacking semantics) to test semantic specificity.
• Neural Processing:
  • Spike sorting was performed to isolate single neurons.
  • Firing rates were calculated in an epoch starting 80ms after word onset, normalized by word duration.
  • Encoding Models: Poisson Generalized Linear Models (GLM) with ridge regression were used to predict spike counts based on:
    • Semantic Embeddings: Word2Vec (static), BERT (contextual, unidirectional), and GPT-2 (contextual, unidirectional).
    • Phonetic Features: IPA phoneme counts.
    • Linguistic Controls: Word duration, frequency, polysemy, syntactic position, and pitch.
• Analysis Techniques:
  • ROC Analysis: To determine how many neurons encode specific words.
  • Population Geometry: Cosine distance calculations between neural response vectors and semantic embedding vectors.
  • Reduced Rank Regression (RRR): To decode semantic embeddings from neural population activity and determine the dimensionality of the neural semantic subspace.
  • Contrastive Coding Analysis: Investigating the relationship between neural and semantic distances for highly similar word pairs.

3. Key Contributions

1. Evidence for Distributed Semantic Coding: Demonstrated that word meanings in the human hippocampus are not encoded by single, narrowly tuned neurons but by distributed populations where individual neurons exhibit "mixed selectivity" (encoding multiple, semantically unrelated words).
2. Contextual Alignment: Showed that neural population activity aligns significantly better with contextual language models (GPT-2, BERT) than with static models (Word2Vec), proving that the hippocampus tracks context-dependent meaning.
3. Discovery of Contrastive Coding: Identified a unique "repulsive" coding mechanism where semantically very similar words (e.g., near-synonyms) evoke more divergent neural patterns than expected, likely to prevent confusion.
4. Frequency-Polysemy Interaction: Revealed that high-frequency words (which are often more polysemous) occupy a distinct geometric region in neural space compared to low-frequency words, driven by their context-dependent variability.

4. Key Results

• Distributed and Mixed Selectivity:
  • Every word in the dataset was encoded by at least 3 neurons, with common words encoded by up to 67 neurons.
  • Individual neurons encoded an average of 6 words, spanning an average of 4 distinct semantic categories.
  • Neurons did not cluster by semantic category; a single neuron could encode words from "family," "objects," and "actions" simultaneously.
• Semantic vs. Phonetic Encoding:
  • Semantic models (using embeddings) explained significantly more variance in neural activity than phonetic models.
  • Adding semantic predictors improved model fit, whereas adding phonetic predictors to a semantic model did not.
  • 57% of neurons showed significantly reduced activity for nonsense (Jabberwocky) words compared to real words, confirming semantic specificity.
• Contextual Embedding Correlation:
  • There was a significant positive correlation between neural distance (dissimilarity in firing patterns) and semantic distance (dissimilarity in embeddings) for contextual models (BERT: $r \approx 0.065$; GPT-2: $r \approx 0.092$).
  • This correlation was negative for static Word2Vec embeddings, suggesting that without context, the brain's representation diverges from static lexical categories.
  • The neural-semantic correlation for GPT-2 reached 76% of the theoretical noise ceiling, indicating a robust signal despite single-neuron variability.
• Contrastive Coding:
  • For the most similar word pairs (semantic distance 0.15–0.35), the correlation between neural and semantic distance turned negative.
  • This implies the brain actively separates highly similar concepts (e.g., "brother" vs. "sister") to avoid confusion, a feature not present in standard LLM embeddings.
• Frequency and Polysemy Effects:
  • High-frequency words (often highly polysemous) showed greater neural distance between their different occurrences compared to low-frequency words.
  • Neural polysemy (variability in neural response to the same word across contexts) strongly correlated with linguistic polysemy ($r = 0.54$).
  • High-frequency words clustered near the center of the neural subspace, making them harder to decode from other categories, while low-frequency words were more distinct.

5. Significance

• Neurocomputational Theory: The study provides a neurocomputational account for how the brain tracks word meaning, supporting the hypothesis that semantics are represented as high-dimensional, distributed population vectors rather than discrete, localized categories.
• Hippocampal Function: It redefines the role of the hippocampus in language, positioning it not just as a memory retrieval center but as a critical node for constructing context-dependent relational meaning in real-time.
• Brain-AI Convergence: The findings suggest a convergence between biological neural coding and artificial intelligence. The brain's use of distributed geometry to represent semantics mirrors the vector space representations in LLMs, though the brain adds a "contrastive" layer to handle fine-grained distinctions that static models miss.
• Methodological Advance: By successfully applying contextual embeddings (GPT-2/BERT) to single-neuron data from naturalistic speech, the paper bridges the gap between computational linguistics and systems neuroscience, offering a new framework for studying language processing in the human brain.