Progress in Natural Language Processing and Predictive Coding

<p>Recently, significant progress has been made in the field of natural language processing: deep learning algorithms are increasingly capable of generating, summarizing, translating, and classifying texts. However, these language models still cannot match human linguistic abilities. The theory of predictive coding offers a preliminary explanation for this discrepancy: while language models are optimized for predicting nearby words, the human brain constantly predicts a hierarchy of representations that spans multiple timescales. To test this hypothesis, we analyzed functional magnetic resonance imaging signals from 304 participants listening to short stories. First, we confirmed that the activations of modern language models linearly map onto the brain's response to speech. Second, we showed that improving these algorithms through predictions that cover multiple timescales enhances brain mapping. Finally, we demonstrated that these predictions are organized hierarchically: the frontoparietal cortices predict higher-level, more distant, and more contextual representations than the temporal cortices. Overall, these results reinforce the role of hierarchical predictive coding in language processing and illustrate how the synergy between neuroscience and artificial intelligence can uncover the computational foundations of human cognition.</p>