Maximum Entropy Loss, the Silver Bullet Targeting Backdoor Attacks in Pre-trained Language Models.
Zhengxiao LiuBowen ShenZheng LinFali WangWeiping WangPublished in: ACL (Findings) (2023)
Keyphrases
- maximum entropy
- language model
- pre trained
- language modeling
- training data
- maximum entropy principle
- document retrieval
- probabilistic model
- speech recognition
- n gram
- information retrieval
- conditional random fields
- retrieval model
- query expansion
- training examples
- test collection
- named entity recognition
- vector space model
- smoothing methods
- control signals
- generative model
- feature vectors
- translation model
- appearance variations