Fig. 3: Layer preferences for embeddings and transformations.
From: Shared functional specialization in transformer-based language models and the human brain
A Layer preferences are visualized on the cortical surface for embeddings (upper) and transformations (lower). While most cortical parcels prefer the final embedding layers, the transformations reveal a cortical hierarchy of increasing layer preference. Only cortical parcels with encoding performance greater than 20% of the noise ceiling for both embeddings and transformations are included for visualization purposes. The same color map for preferred layer is used for both embeddings and transformations. B Partial correlations between brain activity and model-based predictions derived from embeddings (blue) and transformations (red). For each layer, we measured the correlation between transformation-based predictions and brain activity while controlling for the embedding-based predictions (and vice versa). Partial correlations at each layer were averaged across parcels in the cortical language network. Error bars denote 95% bootstrap confidence intervals across subjects (N = 63). C Distribution of the magnitude of layer-to-layer differences in encoding performance for embeddings and transformations; this metric of layer specificity is quantified as the L2 norm of the first differences between encoding performance for neighboring layers. Transformations (red) yield more layer-specific deviations in performance than embeddings (blue). Source data are provided as a Source Data file. Figure made using SUMA, Matplotlib, seaborn, and Inkscape.
