Challenge: a large fraction of attention heads can be randomly pruned with limited effect on accuracy, a new study finds . a second study finds no advantage in pruning attention heads identified to be important based on the location of a head .
Approach: They examine the importance of pruning attention heads on a Transformer-based model . they find no advantage in pruning attention head positions on the BERT model based on location .
Outcome: The results show that pruning strategies on Transformer and BERT models are not important based on location . the results suggest that interpretation of attention heads does not strongly inform pruning strategies.

Similar Papers

Contributions of Transformer Attention Heads in Multi- and Cross-lingual Tasks (2021.acl-long)

Copied to clipboard

Challenge: Prior research has found that only a few attention heads are important in each mono-lingual NLP task and pruning the remaining heads leads to comparable or improved performance of the model.
Approach: They examine the relative importance of attention heads in Transformer-based models to aid their interpretability in cross-lingual and multi-lingual tasks.
Outcome: The proposed model performs better with the remaining heads pruned than with the other models, the authors show .
Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned (P19-1)

Copied to clipboard

Challenge: et al., 2017) show that multi-head attention is important for neural machine translation.
Approach: They evaluate the contribution made by individual attention heads to the overall performance of the Transformer model and analyze the roles played by them in the encoder.
Outcome: The proposed pruning method removes the vast majority of heads without affecting performance.
Differentiable Subset Pruning of Transformer Heads (2021.tacl-1)

Copied to clipboard

Challenge: Recent work shows that a large proportion of the heads in a Transformer’s multi-head attention mechanism can be safely pruned away without significantly harming the performance of the model.
Approach: They propose a method that prunes a Transformer's multi-head attention mechanism away without significantly harming its performance.
Outcome: The proposed method improves on natural language inference and machine translation tasks while offering precise control of sparsity level.
Losing Heads in the Lottery: Pruning Transformer Attention in Neural Machine Translation (2020.emnlp-main)

Copied to clipboard

Challenge: Recent research shows that attention heads are not confident in their decisions and can be pruned.
Approach: They apply the lottery ticket hypothesis to prune heads in early training . they find that the pruned model is 1.5 times faster at inference .
Outcome: The proposed method is 1.5 times faster at inference, but at the cost of longer training.
Revealing the Dark Secrets of BERT (D19-1)

Copied to clipboard

Challenge: Existing models of BERT-based learning systems are lacking specific mechanisms that contribute to its success.
Approach: They propose to use GLUE tasks to analyze the interpretation of self-attention, which is one of the underlying components of BERT.
Outcome: The proposed model outperforms the regular model on GLUE tasks by disabling attention in certain heads.
Telling BERT’s Full Story: from Local Attention to Global Aggregation (2021.eacl-main)

Copied to clipboard

Challenge: Recent work discouraging the use of attention distributions for explaining a model’s behaviour suggests that attention distribution can provide insights into local behaviour of attention heads.
Approach: They propose a distinction between local patterns revealed by attention and global patterns that refer back to the input and analyze BERT from both angles.
Outcome: The proposed model can explain local behaviour of attention heads by comparing local and global patterns from both angles.
Better Explain Transformers by Illuminating Important Information (2024.findings-eacl)

Copied to clipboard

Challenge: Existing explanations focus on the input and output of the Transformers, resulting in confusing results.
Approach: They propose to highlight important information and eliminate irrelevant information by a refined information flow on top of the layer-wise relevance propagation method.
Outcome: The proposed method outperforms baseline models on classification and question-answering datasets with over 3% to 33% improvement on explanation metrics.
On the Distribution, Sparsity, and Inference-time Quantization of Attention Values in Transformers (2021.findings-acl)

Copied to clipboard

Challenge: Recent work shows that attention can be pruned to zeros with minimal loss in accuracy.
Approach: They propose a pruning technique which quantizes attention to a 3-bit format without retraining . they find that 80% of attention values can be pruned to zeros with minimal loss in accuracy .
Outcome: The proposed approach produces only a few unique attention values with minimal loss in accuracy.
ALPS: Attention Localization and Pruning Strategy for Efficient Adaptation of Large Language Models (2025.findings-acl)

Copied to clipboard

Challenge: Prior research has focused on optimizing general-purpose large language models to downstream tasks . however, these approaches inherently introduce data dependency, which hinders generalization and reusability.
Approach: They propose an algorithm that localizes the most task-sensitive attention heads and prunes by restricting attention training updates to these heads, thereby reducing alignment costs.
Outcome: The proposed algorithm achieves 2% performance improvement over baselines on three tasks while localizing the most task-sensitive attention heads.
How Much Does Attention Actually Attend? Questioning the Importance of Attention in Pretrained Transformers (2022.findings-emnlp)

Copied to clipboard

Challenge: Pretrained language models use the attention mechanism to contextualize input inputs . but, we find that it is not as important as thought for pretrained models .
Approach: They propose a probing method that replaces input-dependent attention matrices with constant ones.
Outcome: The proposed method improves performance of pretrained language models without input-dependent attention.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations