下一句会是什么？我们是否高估了预测编码理论？

Core Argument - The article explores the relationship between large language models (LLMs) like ChatGPT and the brain's language processing mechanisms, questioning whether LLMs capture deep cognitive processes or if their predictive capabilities are merely coincidental [1][12]. Group 1: Predictive Coding Theory - Predictive coding theory, proposed by Karl Friston in the 1990s, suggests that the brain actively predicts future events and adjusts its predictions based on sensory input to minimize prediction errors [1][2]. - This theory has gained traction as it provides a coherent framework for understanding various cognitive functions, including language processing, where the brain anticipates upcoming words and sentences [3][4]. Group 2: Neural Network Language Models (NNLM) - NNLMs are artificial neural networks designed for word prediction tasks, leveraging vast amounts of natural language text for training, which allows them to learn statistical patterns across different text types [6][9]. - Recent advancements in NNLMs have led to the development of fine-tuning techniques, enabling models to adapt learned representations for various language tasks, improving performance compared to models trained from scratch [6][10]. Group 3: Neuroscience Research Using NNLM - NNLMs have been employed in neuroscience to predict brain responses to natural language, with studies showing that models based on language representations outperform those using non-contextual embeddings [10][11]. - Research indicates a strong correlation between a model's word prediction accuracy and its ability to explain brain activity, suggesting that word prediction may be fundamental to language processing [10][11]. Group 4: Alternative Explanations - Antonello and Huth challenge the predictive coding theory, proposing that the success of language models may stem from their ability to capture universal information rather than predictive capabilities [12][17]. - Their research indicates that the correlation between model performance and brain response may be due to the generalizability of the representations used, rather than evidence of predictive coding in the brain [12][14]. Group 5: Future Research Directions - Future studies should aim to identify measurable phenomena that can distinctly demonstrate whether the brain employs predictive coding during language processing, potentially providing stronger evidence for or against the theory [18].