Analyzing Redundancy in Pretrained Transformer Models

By Fahim Dalvi, Hassan Sajjad, Nadir Durrani, Yonatan Belinkov. 2020

110 citations recorded by the source corpus.

Transformer-based deep NLP models are trained using hundreds of millions of parameters, limiting their applicability in computationally constrained environments. In this paper, we study the cause of these limitations by defining a notion of Redundancy, which we categorize into two classes: General Redundancy and Task-specific Redundancy. We dissect two popular pretrained models, BERT and XLNet, studying how much redundancy they exhibit at a representation-level and at a more fine-grained neuron-level. Our analysis reveals interesting insights, such as: i) 85% of the neurons across the network are redundant and ii) at least 92% of them can be removed when optimizing towards a downstream task. Based on our analysis, we present an efficient feature-based transfer learning procedure, which maintains 97% performance while using at-most 10% of the original neurons.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Model Architectures, Efficiency & Compression

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Model Architectures

In Efficiency & Compression

Citations in the corpus

Probing Classifiers: Promises, Shortcomings, and Advances (2021, methodology): “motivated an efficient transfer-learning procedure (Dalvi et al. 2020).”
AMMUS : A Survey of Transformer-based Pretrained Models in Natural Language Processing (2021, background): “Pruning-based fine-tuning - Recent studies [164]–[166], [237], [238] show that deep pre-trained language models have redundancy.”
ShortGPT: Layers in Large Language Models are More Redundant Than You Expect (2024, background): “In Dalvi et al. (2020), researchers dissect two pre-trained models, BERT Devlin et al. (2018) and XLNet Yang et al. (2019), studying how much redundancy they exhibit at a representation level and a more fine-grained neuron-level.”
Finding Neurons in a Haystack: Case Studies with Sparse Probing (2023, background): “…individual neurons seemed to have very limited effect, indicating that either the “fact” was computed previously, and these neurons are simply responding to this fact, or that there is some amount of ensembling or redundancy (Dalvi et al., 2020) that is robust to the deletion of a single neuron.”
Probing Classifiers: Promises, Shortcomings, and Alternatives (2021, methodology): “An analysis of redundancy in language models (again using probing classiﬁers and other methods) motivated an efﬁcient transfer-learning procedure (Dalvi et al. 2020).”
Analyzing Individual Neurons in Pre-trained Language Models (2020, unknown): “…Dalvi et al., 2019; Suau et al., 2020) and have other potential beneﬁts such as controlling bias and manipulating system’s behaviour (Bau et al., 2019), model distillation and compression (Rethmeier et al., 2020), efﬁcient feature selection (Dalvi et al., 2020), and guiding architectural search.”