A population code for semantics in human hippocampus

When you listen to someone speak, your brain is constantly tracking not just the sounds of words but their meanings. A new study has found that the hippocampus, a brain region best known for its role in memory, encodes the meaning of words in a surprisingly sophisticated way: through coordinated patterns of activity spread across hundreds of neurons working together, rather than having individual neurons dedicated to specific concepts. Think of it like a musical chord rather than a single note. No one neuron represents 'apple,' but the combined firing pattern of many neurons creates a unique neural signature for each word's meaning in context.

Researchers recorded the activity of hundreds of individual neurons in the hippocampi of epilepsy patients who already had electrodes implanted for medical reasons, while those patients listened to narrative speech. By analyzing these population-level activity patterns, the team found something striking: the 'distance' between the brain's responses to two different words tracked how semantically different those words are in meaning. The brain's coding scheme matched most closely with so-called contextual language models, specifically GPT-2 and BERT, which are artificial intelligence systems that assign different representations to the same word depending on its surrounding context. Non-contextual models, which treat each word as having a fixed meaning regardless of context, did not match the brain's patterns. The researchers also discovered that the brain handles common, everyday words differently from rare ones, likely because common words tend to have multiple meanings depending on context, a property called polysemy. And for words that are very close in meaning, the brain actually pushes their neural representations apart, a strategy the researchers call contrastive coding, which may help prevent confusion between similar concepts.