Scientists recorded individual neurons in bilingual brains for the first time and found that the brain does not translate words, it does something very similar to vector space isomorphism in LLMs instead
SUMMARYResearchers recorded single-neuron activity in the hippocampi of four bilingual people and found that English and Spanish words are arranged in the brain as shared semantic maps rather than direct word-for-word translations. The geometric structure of these maps matched across both languages, and a multilingual BERT model showed a similar cross-language organization. The findings suggest a common representation strategy for meaning in human brains and language models.
I'm sure many of you know about the concept of vector spaces / latent spaces in LLMs. If you have no idea, here is a quick introduction on YouTube.
Now, the reposted article is about a study published in the journal Cell 00579-9?_returnURL=https%3A%2F%2Flinkinghub.elsevier.com%2Fretrieve%2Fpii%2FS0092867426005799%3Fshowall%3Dtrue) in which they "recorded the activity of individual neurons in the hippocampi of four bilingual bilingual people [...] at the level of single neurons, watching in real time as actual cells responded to actual words in two languages simultaneously".
They concluded the following:
Rather than linking individual words through shared cells, the brain organized all the words in each language onto a geometric map. On this map, words were positioned according to their meaning relative to other words. “Dog” and “wolf” sat close together because they are semantically related animals. “Fork” was far from both of them because it belongs to an entirely different conceptual category. The spatial relationships between words on the map reflected the actual relationships between their meanings.
The critical finding was that this map had the same geometric structure in both languages. “Dog” in the English map occupied a position relative to its neighboring concepts that mirrored the position of “perro” in the Spanish map. The concepts were arranged identically. Only the neurons used to read those positions were different.
“It’s like looking into a room from a different window. Everything inside is the same, but the perspective is different,” said senior author Sameer Sheth of Baylor College of Medicine. The brain reads the same conceptual room through language-specific neurons that each provide their own viewing angle.
If this sounds familiar from LLMs, you get confirmation:
The researchers made one additional comparison that adds an unexpected dimension to the finding. They analyzed mBERT, a large language model trained to understand more than 100 languages, and found that it organizes words across languages using the same kind of shared geometric structure that the human hippocampus uses.
The AI model was not designed with any knowledge of how the human brain handles bilingualism. It arrived at the same solution through training on language data alone. The convergence suggests that the shared-geometry approach to multilingual representation may be a deep solution to the problem of handling multiple languages in a single system, one that emerges naturally in both biological and artificial neural networks when they are exposed to enough language across multiple tongues.
What this means is, that the human brain does apparently something very similar to what is known as vector space isomorphism.
Amazing, what a time to be alive!