Predictive Coding or Just Feature Discovery? An Alternative Account of Why Language Models Fit Brain Data

Area of Science:

• Neuroscience
• Computational Linguistics
• Artificial Intelligence

Background:

• Recent studies demonstrate neural network language models (NLMs) accurately predict brain activity during natural language processing.
• A prevailing hypothesis suggests NLMs and the brain share the objective of predicting upcoming words, aligning with predictive coding theories.

Purpose of the Study:

• To investigate the underlying reasons for the effectiveness of NLMs in predicting neural responses.
• To evaluate the hypothesis that word prediction is the primary driver of NLM success in modeling brain activity.

Main Methods:

• Analysis of NLM representations to determine their correlation with brain responses.
• Comparison of representations optimized for word prediction against other representations within NLMs.
• Evaluation of alternative explanations for NLM performance in predicting neural data.

Main Results:

• The ability of NLMs to predict future words does not uniquely or best explain their representational similarity to brain activity.
• Representations within an NLM that excel at word prediction are less effective as brain models compared to other representations.
• Evidence suggests NLMs' success stems from their comprehensive capture of diverse linguistic phenomena.

Conclusions:

• The predictive coding hypothesis, based solely on word prediction, may not fully account for NLM effectiveness in neuroscience.
• Alternative explanations, such as the broad linguistic feature representation within NLMs, offer a more compelling account for their success in predicting brain responses.
• Future research should explore the specific linguistic features captured by NLMs that contribute to their predictive power in neural modeling.