%Aigaion2 BibTeX export from Idiap Publications
%Saturday 21 December 2024 04:58:44 PM

@ARTICLE{Fehr_NVRegularisation_2023,
         author = {Fehr, Fabio and Henderson, James},
       projects = {Idiap, EVOLANG},
          title = {Nonparametric Variational Regularisation of Pretrained Transformers},
        journal = {ArXiv},
           year = {2023},
            url = {https://arxiv.org/abs/2312.00662},
            doi = {https://doi.org/10.48550/arXiv.2312.00662},
       abstract = {The current paradigm of large-scale pre-training and fine-tuning Transformer large language models has lead to significant improvements across the board in natural language processing. However, such large models are susceptible to overfitting to their training data, and as a result the models perform poorly when the domain changes. Also, due to the model's scale, the cost of fine-tuning the model to the new domain is large. Nonparametric Variational Information Bottleneck (NVIB) has been proposed as a regulariser for training cross-attention in Transformers, potentially addressing the overfitting problem. We extend the NVIB framework to replace all types of attention functions in Transformers, and show that existing pretrained Transformers can be reinterpreted as Nonparametric Variational (NV) models using a proposed identity initialisation. We then show that changing the initialisation introduces a novel, information-theoretic post-training regularisation in the attention mechanism, which improves out-of-domain generalisation without any training. This success supports the hypothesis that pretrained Transformers are implicitly NV Bayesian models.}
}