IC-1410FF's contextualization effects in BERT and GPT-2 are largely canceled by the residual connection and layer normalization, with LN's γ weights specifically shrinking the outlier dimensions in FF output
The paper shows that the contextualization change introduced by FF is progressively undone by the subsequent residual connection and layer normalization. The L2 norm of the res2 bypass vector is more than twice that of the FF output in most layers, so the original representation dominates. Additionally, the FF output has outlier values in specific dimensions, and the LN weight γ is negatively correlated with these dimensions (Pearson r from -0.45 to -0.78 in BERTs, -0.22 to -0.59 in GPT-2). When the bottom 1% of dimensions with the lowest γ are excluded from the norm calculation, the FF contextualization change drops from 0.21 to 0.09 in BERT and from 0.15 to 0.02 in GPT-2, indicating that FF's effect is realized through very specific dimensions that LN erases. This suggests redundancy in the transformer layer's processing.
Evidence
correlational
Key metric
res2 bypass vector L2 norm >2x FF output in most layers; Pearson r between LN γ and mean absolute FF output by dimension: -0.45 to -0.78 (BERTs), -0.22 to -0.59 (GPT-2); FF contextualization change with bottom 1% γ dimensions ignored: 0.21→0.09 (BERT), 0.15→0.02 (GPT-2)
Caveat
The cancellation is weakened in the 10th-11th layers of BERT and former layers of GPT-2 where FF's contextualization is relatively large; in RoBERTa the Pearson r is small (-0.02 to -0.28) though Spearman ρ is large (-0.46 to -0.56), suggesting the mechanism may differ.