Outliers Dimensions that Disrupt Transformers Are Driven by Frequency
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2022
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866914836823146496 |
|---|---|
| author | Puccetti, Giovanni Rogers, Anna Drozd, Aleksandr Dell'Orletta, Felice |
| author_facet | Puccetti, Giovanni Rogers, Anna Drozd, Aleksandr Dell'Orletta, Felice |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2205_11380 |
| institution | arXiv |
| publishDate | 2022 |
| record_format | arxiv |
| spellingShingle | Outliers Dimensions that Disrupt Transformers Are Driven by Frequency Puccetti, Giovanni Rogers, Anna Drozd, Aleksandr Dell'Orletta, Felice Computation and Language Artificial Intelligence 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. |
| title | Outliers Dimensions that Disrupt Transformers Are Driven by Frequency |
| topic | Computation and Language Artificial Intelligence |
| url | https://arxiv.org/abs/2205.11380 |