Surprisal from Larger Transformer-based Language Models Predicts fMRI Data More Poorly

Technical Summary: Surprisal from Larger Transformer-based Language Models Predicts fMRI Data More Poorly

Problem Statement
Expectation-based theories of syntactic comprehension posit that processing difficulty is proportional to information-theoretic surprisal (the negative log probability of a word given its context). While Transformer-based language models (LMs) have been widely adopted to estimate surprisal, recent studies have identified an "inverse scaling" phenomenon on latency-based data (reading times): as LMs increase in size (parameter count) and training data, their surprisal estimates become less predictive of human reading times. Larger models, which achieve lower perplexity and better predict rare tokens, assign lower surprisal values that diverge from human behavioral data.

However, previous investigations into this phenomenon using brain imaging data (fMRI) yielded inconclusive results. A study by Oh et al. (2022) using a relatively small set of LMs found no clear scaling trend on fMRI data, leaving open the possibility that inverse scaling is specific to latency-based measures. Furthermore, other research using raw LM vector representations (rather than surprisal) on fMRI data has shown positive scaling, suggesting brain imaging data might behave differently. This study addresses whether the inverse scaling relationship holds for fMRI data when evaluated with a comprehensive set of models.

Methodology
The authors conducted a comprehensive evaluation using surprisal estimates from 17 pre-trained, autoregressive Transformer-based LMs across three model families: GPT-2, GPT-Neo, and OPT. These models varied significantly in parameter count. The study utilized two distinct functional magnetic resonance imaging (fMRI) datasets:

1. Natural Stories fMRI (Shain et al., 2020): Time-series Blood Oxygenation Level-Dependent (BOLD) signals recorded from 78 subjects listening to naturalistic English stories.
2. Pereira fMRI (Pereira et al., 2018): Per-sentence BOLD signals aggregated across language network regions from subjects reading English passages.

Experimental Procedure:

• Surprisal Calculation: Texts were tokenized using model-specific byte-pair encoding. Word-level surprisal was calculated by aggregating token-level estimates. To address tokenizer artifacts where leading whitespace probabilities sum to greater than one, the study incorporated whitespace-trailing decoding, reallocating leading whitespace probability to the preceding word.
• Signal Alignment:
  • For Natural Stories fMRI, word-level surprisal estimates were convolved with a canonical hemodynamic response function (HRF) to align with the continuous BOLD time series.
  • For Pereira fMRI, where responses were per-sentence, the surprisal estimate of the sentence-final word was used as the predictor, as the preceding context is fully accounted for in that estimate.
• Regression Modeling: Linear Mixed-Effect (LME) models were fitted to predict BOLD signals.
  • Predictors included the surprisal estimate, word position, and word length (or sentence length for Pereira).
  • Models included random effects for subjects and stories/sentences.
  • Model fit was evaluated using log-likelihood on held-out test sets.

Key Results
The study found a statistically significant inverse scaling relationship between model size and predictive power on both fMRI datasets:

• Natural Stories fMRI: As the parameter count of the LMs increased, the log-likelihood of the regression models predicting BOLD signals significantly decreased ($p < 0.01$). This resolves the inconclusive findings of Oh et al. (2022) by utilizing a broader range of model sizes.
• Pereira fMRI: A similar significant decrease in predictive power was observed as model size increased ($p < 0.05$).

The results indicate that larger LMs, despite having higher per-word estimated probabilities (lower perplexity), provide a poorer fit to human neural data than smaller LMs.

Significance and Claims
The primary contribution of this work is the demonstration that the inverse scaling phenomenon of surprisal estimates is not specific to latency-based measures (like reading times) but also applies to neural data (fMRI).

• Resolution of Inconclusive Data: The study clarifies previous ambiguity regarding fMRI data, showing that the trend observed in behavioral data generalizes to brain imaging when a sufficiently diverse set of models is evaluated.
• Nature of the Trend: The authors note that while the inverse relationship exists on fMRI data, it is "not as strict" as on latency-based data. They speculate this is due to the higher noise levels inherent in fMRI measurements compared to behavioral latency data.
• Implications for Research: The findings suggest that brain imaging data does not behave substantially differently from latency data in this context. Consequently, the divergence of larger LMs from human processing patterns (likely driven by their improved prediction of rare tokens) is a general phenomenon across psychometric measures. The authors suggest that smaller models may be more appropriate for researchers seeking to understand human sentence processing mechanisms using LMs, as larger models may overfit to statistical regularities in training data that do not align with human cognitive constraints.

Limitations
The authors acknowledge that the models and datasets are English-centric, and the findings may not replicate in other languages. Additionally, the study relies on published fMRI datasets, and the inverse scaling on the Pereira dataset was less statistically robust than on the Natural Stories dataset, potentially due to data sparsity when averaging across subjects.