Besides the awesome neuroscience I learned at HSN and @sfn.org, I was reminded of my incredible support system and the many inspiring women in neuroscience who currently guide and influence me! @nicolecrust.bsky.social @libertysays.bsky.social @storiesofwin.bsky.social to name a few.

If you're attending HSN or #SfN2025 #SfN25 be sure to come by my poster to learn about my new work, "Neural signatures of impaired semantic contextualization in autism."

SFN info: Board KK2, Sun Nov 16. 13-1700PM

Really looking forward to presenting my new research on semantic coding in single neurons in autism this Sunday at the Society for Neurobiology of Language conference! Be sure to check out all the presentations from the Hayden lab! @benhayden.bsky.social @snlmtg.bsky.social

Similarly, a neuron may be selective to a specific meaning of a word, and therefore only fire when the word is used in that context, resulting in sparse coding.

Contextual LLMs and the brain take this into account. Since neurons and LLMs are contextual coders, their distances correlate with word polysemy, causing the greatest changes in neural activity between high frequency and thus more polysemous words (such as function words, actions, and pronouns).

How does the brain represent different occurrences of the same word? Neural activity varies greatly for different instances of the same word. Words that are more frequently used in the English language are more polysemous, meaning they can mean many different things, depending on their context.

reflecting a negative correlation to LLMs only for very similar word pairs. This contrastive coding likely helps prevent confusion for different words with very similar meanings.

After comparing the angles between neural vectors and model vectors for all word pairs, we found that the neural encoding of semantics is highly contextual and positively correlates only with LLMs that have contextual embeddings (GPT, BERT).

We wanted to know if the brain also employed these coding principles. For each word, its neural embedding vector comprised the firing rate response of each neuron for that word.

When considering words related to the concept of family, of the 28 neurons encoding "father" and 19 encoding "dad," only 7 neurons encode both words. Each neuron encoded a median of 6 words and their words spanned a median of 4 semantic categories (out of 11).

‘Hit’ was the most densely coded word at 67 neurons and ‘was’ and ‘and’ were the most sparse at 3 neurons. Even words coding specific concepts like ‘God’ were densely coded (60 neurons).

We recorded the activity of hundreds of neurons in the human hippocampus while people listened to stories. We found a dense and distributed code for semantics based on comparing each neuron's firing rate across words. A single word was encoded by many neurons and each neuron encoded many words.

I am incredibly proud to share my first, first-author paper as a postdoc with @benhayden.bsky.social . How does the human hippocampus, known for encoding concepts, represent the meanings of words while listening to narrative speech?
www.biorxiv.org/content/10.1...

Highly recommend this book for learning how to be a more empathetic, compassionate, and meaningful person in your community and workplace, especially the lab! Ruchika provides practical, actionable strategies to enhance inclusivity and belonging in science.

Finally cold enough to wear a cowl in Houston! ❄️☃️