Language: In search of neural code

Item: Language: In search of neural code
Rating: 4
Author: PostSingularity Institute

by Jean-Rémy KING

Life & Natural Systems Computing & Cybersecurity UComputing and Information Technology UYComputer science

🎙 Jean-Rémy KING 👥 305 📅 March 19, 2026 ⏱ 86 min 👁 99 📄 original study 🧭 2026-08-16

Available in: English (current) Français

Keywords

languagebraindeep learningencoding modelsneuroAI

Summary

Jean-Rémy KING presents a seminar on using deep learning models to understand how language is processed in the human brain. The approach involves comparing activations of artificial neural networks (like wav2vec2 and LLMs) to brain responses recorded with fMRI, MEG, and intracranial EEG. Encoding models, based on linear regression, reveal a hierarchy of representations in the brain that aligns with the layers of the AI models. The similarity increases with model size and context, and middle layers are most similar to brain activity. The method allows for transfer across languages and modalities (listening/reading). A study with children (2-12 years) implanted with electrodes shows that brain representations of language develop with age, and their similarity to LLM representations increases. The speaker discusses the concept of convergence and acknowledges limitations, but sees this as an unprecedented opportunity to model language acquisition and processing. The presentation includes Q&A sessions and references several papers.

150 words

Critical Evaluation

Value of the Information & Strength of the Argument

The presentation provides valuable insights into the alignment between deep learning models and brain activity, supported by multiple studies and clear explanations of the methodology. The argumentation is solid, with careful consideration of limitations and alternative interpretations. The speaker demonstrates a deep understanding of both neuroscience and AI, and the discussion with the audience adds depth to the presentation.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is high, with references to peer-reviewed and preprint studies. The sources are credible, and the methodology is well-described. The title accurately reflects the content. The presentation is a seminar, so it is not a formal publication, but it is based on original research. The speaker acknowledges limitations and discusses potential biases.

128 words

Title / Content Match

The title accurately reflects the content: the search for the neural code of language through comparisons with deep learning models.

Quality & Reliability

8/10

Presentation by a CNRS researcher at ENS and Meta AI, based on peer-reviewed and preprint studies, with rigorous methodology (encoding models, comparisons across modalities and ages). Limitations are acknowledged. The content is technical and detailed, but the video is a seminar recording with limited production quality.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Evanson, L., et al. (2025). Emergence of language in the developing brain. arXiv. — Referenced as a study on language development in children.
Lévy, J., et al. (2025). Brain-to-text decoding: A non-invasive approach via typing. arXiv preprint arXiv:2502.17480. — Referenced as a study on decoding brain activity to text.
Banville, H., et al. (2025). Scaling laws for decoding images from brain activity. arXiv preprint arXiv:2501.15322. — Referenced as a study on scaling laws for brain decoding.

Concurring Sources

Schrimpf, M., et al. (2021). The neural architecture of language: Integrative modeling converges on predictive processing. PNAS. — Similar findings that language models align with brain activity.

Dissenting Sources

Bowers, J. S., et al. (2023). Deep problems with neural network models of human vision. Behavioral and Brain Sciences. — Raises concerns about the biological plausibility of deep learning models as models of human cognition.

Contribution & Novelties

The presentation offers a comprehensive overview of recent work in neuroAI, demonstrating that deep learning models can serve as powerful tools to understand the neural basis of language. The key novelty is the systematic comparison of model layers and brain regions, revealing a convergence of representations. The inclusion of developmental data from children is particularly novel. The speaker also discusses the concept of convergence and its implications.

Pour aller plus loin :

Encoding models in neuroscience — Provides background on the methodology used.
Large language models — Context on the AI models used.
Convergent evolution — The concept discussed in the presentation.

102 words

Radar Profile

The radar profile shows high scores in information quantity, quality, and technical level, with a slightly lower but still high reliability score. This indicates a dense, technically rigorous presentation with strong scientific backing.

Reliability 8/10