Researchers including a team at Princeton University recorded directly from the brains of nine patients using implanted electrodes while each listened to a 30-minute podcast, comparing the recordings against a modern language-prediction model.
The brain activity matched three specific behaviors of autoregressive language models: it generated a continuous prediction of the upcoming word before that word was heard, it compared that prediction against the actual word to register surprise, and it represented each word using the same kind of context-dependent encoding the models use.
The overlap suggests the brain and modern language models, despite being built from utterly different material, converge on some of the same computational strategies for processing ongoing speech. The result came from just nine patients undergoing electrode monitoring for unrelated epilepsy treatment, a specialized and small sample, so replicating the exact match across more listeners and languages is the next step.
Read the original study
Goldstein et al., Nature Neuroscience, 2022 · doi.org/10.1038/s41593-022-01026-4
Two very different machines, one shared trick.
Source
- Goldstein, A., Zada, Z., Buchnik, E., Schain, M., Price, A., Aubrey, B., Nastase, S. A., Feder, A., Emanuel, D., Cohen, A., Jansen, A., Gazula, H., Choe, G., Rao, A., Kim, C., Casto, C., Fanda, L., Doyle, W., Friedman, D., . . . Hasson, U. (2022). Shared computational principles for language processing in humans and deep language models. Nature Neuroscience, 25(3), 369–380. https://doi.org/10.1038/s41593-022-01026-4