BERT Busters: Outlier Dimensions that Disrupt Transformers
원문 초록
Multiple studies have shown that Transformers are remarkably robust to pruning. Contrary to this received wisdom, we demonstrate that pre-trained Transformer encoders are surprisingly fragile to the removal of a very small number of features in the layer outputs (<0.0001% of model weights). In case of BERT and other pre-trained encoder Transformers, the affected component is the scaling factors and biases in the LayerNorm. The outliers are high-magnitude normalization parameters that emerge early in pre-training and show up consistently in the same dimensional position throughout the model. We show that disabling them significantly degrades both the MLM loss and the downstream task performance. This effect is observed across several BERT-family models and other popular pre-trained Transformer architectures, including BART, XLNet and ELECTRA; we also show a similar effect in GPT-2.
함께 살펴볼 논문
공통 태그를 근거로 연결했습니다. 실제 인용 관계나 논문의 우수성을 뜻하지 않습니다.
Car Object Counting and Position Estimation via Extension of the CLIP-EBC Framework
SeqPE: Transformer with Sequential Position Encoding
We Should Identify and Mitigate Third-Party Safety Risks in MCP-Powered Agent Systems
JoFormer (Journey-based Transformer): Theory and Empirical Analysis on the Tiny Shakespeare Dataset
POSS: Position Specialist Generates Better Draft for Speculative Decoding
ComRoPE: Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices
토론
이 논문과 연결된 포럼 글이 아직 없습니다.