Outlier Dimensions that Disrupt Transformers are Driven by Frequency
Research output: Contribution to conference › Paper › Research › peer-review
While Transformer-based language models are generally very robust to pruning, there is the recently discovered outlier phenomenon: disabling only 48 out of 110M parameters in BERT-base drops its performance by nearly 30% on MNLI. We replicate the original evidence for the outlier phenomenon and we link it to the geometry of the embedding space. We find that in both BERT and RoBERTa the magnitude of hidden state coefficients corresponding to outlier dimensions correlates with the frequency of encoded tokens in pre-training data, and it also contributes to the “vertical” self-attention pattern enabling the model to focus on the special tokens. This explains the drop in performance from disabling the outliers, and it suggests that to decrease anisotropicity in future models we need pre-training schemas that would better take into account the skewed token distributions.
Original language | English |
---|---|
Publication date | 2022 |
Number of pages | 19 |
Publication status | Published - 2022 |
Event | 2022 Findings of the Association for Computational Linguistics: EMNLP 2022 - Abu Dhabi, United Arab Emirates Duration: 7 Dec 2022 → 11 Dec 2022 |
Conference
Conference | 2022 Findings of the Association for Computational Linguistics: EMNLP 2022 |
---|---|
Country | United Arab Emirates |
City | Abu Dhabi |
Period | 07/12/2022 → 11/12/2022 |
Bibliographical note
Publisher Copyright:
© 2022 Association for Computational Linguistics.
ID: 347300129