| Challenge: | Existing methods to accelerate pretraining of transformer-based models are computationally expensive and degrade performance on downstream tasks. |
| Approach: | They propose a "token dropping" method to accelerate the pretraining of transformer-based models by 25% . they leverage the already built-in masked language modeling loss to identify unimportant tokens with practically no computational overhead. |
| Outcome: | The proposed method reduces the pretraining cost of BERT models by 25% while achieving similar overall performance on downstream tasks. |
Similar Papers
Revisiting Token Dropping Strategy in Efficient BERT Pretraining (2023.acl-long)
Copied to clipboard
| Challenge: | Token dropping is a recently-proposed strategy to speed up the pretraining of masked language models, such as BERT. |
| Approach: | They propose a semantic-consistent learning method to improve token dropping by skipping the computation of a subset of input tokens at several middle layers. |
| Outcome: | The proposed method achieves consistent and significant performance gains across all tasks and model sizes. |
TR-BERT: Dynamic Token Reduction for Accelerating BERT Inference (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing pre-trained language models (PLMs) are expensive in inference, making them impractical in resource-limited real-world applications. |
| Approach: | They propose a dynamic token reduction approach to accelerate PLMs' inference by adapting the layer number of each token to avoid redundant calculation. |
| Outcome: | The proposed approach speeds up BERT by 2-5 times and improves performance in long-text tasks with less computation. |
Frustratingly Simple Pretraining Alternatives to Masked Language Modeling (2021.emnlp-main)
Copied to clipboard
| Challenge: | Masked language modeling (MLM) is widely used in natural language processing for self-supervised learning of text representations. |
| Approach: | They propose to use token-level classification tasks as main pretraining objectives instead of Masked language modeling (MLM) . Empirical results show that pretraining a model with 41% of the BERT-BASE’s parameters, BERT MEDIUM results in only a 1% drop in GLUE scores with their best objective. |
| Outcome: | Empirical results show that the proposed methods achieve comparable or better performance to MLM using a BERT-BASE architecture. |
Pyramid-BERT: Reducing Complexity via Successive Core-set based Token Selection (2022.acl-long)
Copied to clipboard
| Challenge: | Existing models that use heuristics to shorten sequence lengths are computationally prohibitive. |
| Approach: | They propose a new method to shorten sequence lengths by transforming tokens through encoders and a core-set based token selection method that avoids expensive pre-training and fine tuning. |
| Outcome: | The proposed model outperforms existing models on GLUE benchmarks and Long Range Arena datasets and demonstrates that it is cost-effective and space-efficient. |
Mask More and Mask Later: Efficient Pre-training of Masked Language Models by Disentangling the [MASK] Token (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Large-scale pre-trained MLMs can be used to generalize well to a wide range of tasks. |
| Approach: | They propose to append [MASK]s at a later layer to reduce sequence length for earlier layers. |
| Outcome: | The proposed method outperforms RoBERTa for 6 out of 8 GLUE tasks on average by 0.4%. |
Focus on the Core: Efficient Attention via Pruned Token Compression for Document Classification (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Pre-trained transformers suffer from a computationally expensive self-attention mechanism that interacts with all tokens, including those unfavorable to classification performance. |
| Approach: | They propose to integrate token pruning and token combining strategies to improve model performance and reduce computational demands. |
| Outcome: | Experiments with various datasets show that the proposed model performs better than baseline models, with the best improvement over the existing model. |
GhostBERT: Generate More Features with Cheap Operations for BERT (2021.acl-long)
Copied to clipboard
| Challenge: | Existing studies show that some parameters in pre-trained language models can be pruned away without severe accuracy degradation. |
| Approach: | They propose a method which generates more features with very cheap operations from the remaining features and can be applied to unpruned BERT models to enhance their performance. |
| Outcome: | Empirical results on the GLUE benchmark on three backbone models (i.e., BERT, RoBERTa and ELECTRA) verify the efficacy of the proposed method. |
AdapLeR: Speeding up Inference by Adaptive Length Reduction (2022.acl-long)
Copied to clipboard
| Challenge: | Pre-trained language models have shown stellar performance in downstream tasks, but their excessive computational costs and high latency hinder their usage in resource-limited settings. |
| Approach: | They propose a method that dynamically eliminates less contributing tokens through layers, resulting in shorter lengths and consequently lower computational cost. |
| Outcome: | The proposed method shows speedups up to 22x during inference time without much sacrifice in performance. |
Effective Pretraining Objectives for Transformer-based Autoencoders (2022.findings-emnlp)
Copied to clipboard
| Challenge: | ELECTRA is more accurate than BERT, but it is not clear if this is due to its innovative architecture or to the long and extensive training, which highly increases the computation cost for obtaining the final language model. |
| Approach: | They propose to replace BERT’s Masked Language Modeling objective (MLM) with Token Detection (TD) by using a statistical approach to generate light tokens. |
| Outcome: | The proposed method can replace ELECTRA's computationally heavy generators without a significant drop in performance. |
DeeBERT: Dynamic Early Exiting for Accelerating BERT Inference (2020.acl-main)
Copied to clipboard
| Challenge: | Large-scale pre-trained language models such as BERT are notorious for being slow in both training and inference. |
| Approach: | They propose a method to accelerate BERT inference by inserting extra classification layers between each transformer layer of BERT. |
| Outcome: | The proposed method saves up to 40% inference time with minimal degradation in model quality. |