Effects of Parameter Norm Growth During Transformer Training: Inductive Bias from Gradient Descent
Open accessThe DOI matches its Crossref record; 9 of the 9 references it cites were found in a registry or an index.
Abstract
The capacity of neural networks like the widely adopted transformer is known to be very high.Evidence is emerging that they learn successfully due to inductive bias in the training routine, typically a variant of gradient descent (GD).To better understand this bias, we study the tendency for transformer parameters to grow in magnitude (ℓ 2 norm) during training, and its implications for the emergent representations within self attention layers.Empirically, we document norm growth in the training of transformer language models, including T5 during its pretraining.As the parameters grow in magnitude, we prove that the network approximates a discretized network with saturated activation functions.Such "saturated" networks are known to have a reduced capacity compared to the full network family that can be described in terms of formal languages and automata.Our results suggest saturation is a new characterization of an inductive bias implicit in GD of particular interest for NLP.We leverage the emergent discrete structure in a saturated transformer to analyze the role of different attention heads, finding that some focus locally on a small number of positions, while other heads compute global averages, allowing counting.We believe understanding the interplay between these two capabilities may shed further light on the structure of computation within large transformers.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325
Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., & Liu, P. J. (2019) · arXiv · doi:10.48550/arxiv.1910.10683 · Cited by 10,929
A primer in BERTology: What we know about how BERT works
Rogers, A., Kovaleva, O., & Rumshisky, A. (2020) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00349 · Cited by 1,552 · cited 1×
“Despite various attempts to study these biases in transformers (Rogers et al., 2020; Lovering et al., 2021), it remains an interesting open question what they are, or even how to characterize them in a way relevant to the domain of language.”
BackgroundOn the computational power of neural nets
Siegelmann, H. T., & Sontag, E. D. (1992) · Proceedings of the fifth annual workshop on Computational learning theory · doi:10.1145/130385.130432 · Cited by 238
On the Practical Computational Power of Finite Precision RNNs for Language Recognition
Weiss, G., Goldberg, Y., & Yahav, E. (2018) · Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) · doi:10.18653/v1/p18-2117 · Cited by 226
Self-citationImplicit Regularization in Matrix Factorization
Gunasekar, S., Woodworth, B., Bhojanapalli, S., Neyshabur, B., & Srebro, N. (2018) · 2018 Information Theory and Applications Workshop (ITA) · doi:10.1109/ita.2018.8503198 · Cited by 170
Language Models with Transformers
Wang, C., Li, M., & Smola, A. J. (2019) · arXiv · doi:10.48550/arxiv.1904.09408 · Cited by 161
On the Ability and Limitations of Transformers to Recognize Formal Languages
Bhattamishra, S., Ahuja, K., & Goyal, N. (2020) · Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2020.emnlp-main.576 · Cited by 48 · cited 3×
“Past work (Merrill, 2019; Bhattamishra et al., 2020) reveals that saturation permits two useful types of attention heads within a transformer: one that locally targets a small number of positions, and one that attends uniformly over the full sequence, enabling an “average” operation.”
BackgroundOn the Linguistic Capacity of Real-Time Counter Automata
Merrill, W. (2020) · arXiv · doi:10.48550/arxiv.2004.06866 · Cited by 12
Self-citation
Cited by
Overcoming a Theoretical Limitation of Self-Attention
Chiang, D., & Cholak, P. (2022) · Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/2022.acl-long.527 · Cited by 39
Saturated Transformers are Constant-Depth Threshold Circuits
Merrill, W., Sabharwal, A., & Smith, N. A. (2022) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00493 · Cited by 25
“One path toward this goal is via formal analysis of specific network architectures (Merrill, 2021); for example, recurrent neural networks (RNNs).”
BackgroundResultInfluentialSelf-citationThe Parallelism Tradeoff: Limitations of Log-Precision Transformers
Merrill, W., & Sabharwal, A. (2023) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00562 · Cited by 22
“We imagine f ( h , h ) is computed via saturated attention (cf. Merrill et al., 2022), which provides a simple model of the types of attention we can expect to be learned in trans-formers (Merrill et al., 2021).”
MethodSelf-citation(2023) · doi:10.18653/v1/2023.acl-short.38 · Cited by 12
The expressive power of transformers with chain of thought
Merrill, W., & Sabharwal, A. (2023) · arXiv · doi:10.48550/arxiv.2310.07923 · Cited by 4
Self-citation
Metadata
- Authors
- Merrill, W., Ramanujan, V., Goldberg, Y., Schwartz, R., & Smith, N. A.
- Venue
- Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing
- Published
- 2021
- Pages
- 1766-1781
- OpenAlex
- W3200987621
- Cited by
- 13
- References
- 9
- FWCI
- 1.48
- Type
- conference paper
- Language
- en
- Open access
- hybrid
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Effects of Parameter Norm Growth During Transformer Training: Inductive Bias from Gradient Descent. https://app.cochat.ai/papers/10.18653/v1/2021.emnlp-main.133
- The paper
- Merrill, W., Ramanujan, V., Goldberg, Y., Schwartz, R., & Smith, N. A. (2021). Effects of Parameter Norm Growth During Transformer Training: Inductive Bias from Gradient Descent. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. https://doi.org/10.18653/v1/2021.emnlp-main.133
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.