Recent Discovery in Neuroscience: Language encoding in neurons at a single cellular level
A recently published Nature article (read it here) presented an interesting discovery: mapping between the processing of words and sentences to specific neurons in the human brain. More specifically, they mapped the semantic and linguistic processing of words and sentences to not just specific regions or subregions of the brain, but to individual neurons (and a network or ensembles of neurons, by extension).
This is an important discovery, as until now, scientists were only able to map linguistic processes to specific regions and subregions of the brain. For example, it was known already that semantic processing of the words and their (resp.) sentence comprehension occurs in subregions of the temporal and frontal regions. However, the granularity was known only to specific subregions, but not a single-neuron level, that is, until now.
The researchers tested 10 human participants by playing audio recordings consisting of semantically diverse sentences. In total, each human participant listened to approx. 118-144 sentences, consisting of approx. 435-483 words in each. The specific sentences given to each participant were randomly selected (randomized control trial). Additionally, the participants were presented with unstructured word lists, nonwords, and stories.
They collected brain data using custom tungsten microelectrode arrays and silicon Neuropixel arrays (for cross-validation).
On their data processing side, they used word embeddings--a technique from Natural Language Processing, that assigns different weights to the same word occurring in different contexts/sentences--and then applied clustering. They found nine clusters, each was found to be associated with specific contexts (they call them semantic domains).
What did they find? (1) They found high selectivity in semantic and linguistic processing at the single-neuron level. They found that specific neurons only fire (or fire at an increased rate) only for specific word meanings, with 84% of these neurons only firing for their corresponding word meanings.
(2) They were able to predict the semantic domains for words and new words that were not part of the original training dataset by looking at the selective firing rates among the neurons i.e. they checked which specific neurons fired for those words and then remapped those words to their respective domains.
(3) They found that the selectivity in single-neuronal firings for words were dependent on their context. In other words, a specific neuron that fired selectively for a specific word did not fire at the same rate when the word appeared in a context-less list as opposed to a meaningful sentence.
(4) They found that different subnetworks of neurons (which they call cell populations) fire selectively for specific contexts i.e. for specific (meaningful) sentences. They took an interesting approach to make finding. They first concatenated the model weights (from the previous step) from each neuron to create a 133-by-embedding dimension matrix (note: 133 is the number of cell population). Following that, they used distance/similarity measurements between word embeddings (hence their contextual meanings) and the neuronal activity/firing rates to confirm selectivity among neuronal ensembles and specific contexts.