We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.CL

Change to browse by:

cs

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Computation and Language

Title: TiltedBERT: Resource Adjustable Version of BERT

Abstract: In this paper, a novel adjustable fine-tuning method is proposed that improves the training and inference time of the BERT model on downstream tasks. In the proposed method, first, the more important word vectors are detected in each layer by the proposed Attention Context Contribution (ACC) metric. Second, the less important ones are eliminated with the proposed strategy. In the TiltedBERT method, the word vector elimination rate in each layer is controlled by the Tilt-Rate hyper-parameter, and the model learns to work with a considerably lower number of Floating Point Operations (FLOPs) than the original BERTbase model. The proposed method does not need any extra training steps, and also it can be generalized to other transformer-based models. The extensive experiments show that the word vectors in higher layers have less contribution that can be eliminated and improve the training and inference time. Experimental results on extensive sentiment analysis, classification and regression datasets, and benchmarks like IMDB and GLUE showed that the TiltedBERT is effective in various datasets. TiltedBERT improves the inference time of BERTbase up to 5.3 times with less than 0.85% accuracy drop on average. After the fine-tuning by the offline-tuning property, the inference time of the model can be adjusted for a wide range of Tilt-Rate selection. Also, A mathematical speedup analysis is proposed to estimate the TiltedBERT methods speedup accurately. With the help of this analysis, the proper Tilt-Rate value can be selected before finetuning and during offline-tuning phases.
Subjects: Computation and Language (cs.CL)
Cite as: arXiv:2201.03327 [cs.CL]
  (or arXiv:2201.03327v4 [cs.CL] for this version)

Submission history

From: Sajjad Kachuee [view email]
[v1] Mon, 10 Jan 2022 13:04:39 GMT (413kb,D)
[v2] Fri, 14 Jan 2022 14:45:24 GMT (487kb,D)
[v3] Fri, 11 Feb 2022 07:53:45 GMT (581kb,D)
[v4] Thu, 17 Mar 2022 14:04:59 GMT (578kb,D)

Link back to: arXiv, form interface, contact.