mlsanityData to 14 Oct 2026

All Roads Lead to Rome? Exploring the Invariance of Transformers' Representations

arXiv 2305.14555 ↗First posted 23 May 2023cs.CL cs.AI cs.LGcited by 6

Not on the map. Why?

Abstract (arXiv)

Transformer models bring propelling advances in various NLP tasks, thus inducing lots of interpretability research on the learned representations of the models. However, we raise a fundamental question regarding the reliability of the representations. Specifically, we investigate whether transformers learn essentially isomorphic representation spaces, or those that are sensitive to the random seeds in their pretraining process. In this work, we formulate the Bijection Hypothesis, which suggests the use of bijective methods to align different models' representation spaces. We propose a model based on invertible neural networks, BERT-INN, to learn the bijection more effectively than other existing bijective methods such as the canonical correlation analysis (CCA). We show the advantage of BERT-INN both theoretically and through extensive experiments, and apply it to align the reproduced BERT embeddings to draw insights that are meaningful to the interpretability research. Our code is at https://github.com/twinkle0331/BERT-similarity.

Key facts. Cited by 6 papers in the map (data through 14 Oct 2026). Builds on 29 of its 40 references that are in the map. First posted on arXiv 23 May 2023.

Builds on

29 of 40 references.

References · 29

  1. 2017Attention Is All You Needcited by 60,361
  2. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,399
  3. 2020Language Models are Few-Shot Learnerscited by 32,419
  4. 2023GPT-4 Technical Reportcited by 18,396
  5. 2022Training language models to follow instructions with human feedbackcited by 15,971
  6. 2019RoBERTa: A Robustly Optimized BERT Pretraining Approachcited by 13,372
  7. 2018GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understandingcited by 4,899
  8. 2016Gaussian Error Linear Units (GELUs)cited by 3,502
  9. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  10. 2017A Broad-Coverage Challenge Corpus for Sentence Understanding through Inferencecited by 2,682
  11. 2016Density estimation using Real NVPcited by 2,042
  12. 2019BERT Rediscovers the Classical NLP Pipelinecited by 1,059
17 more
  1. 2019What Does BERT Look At? An Analysis of BERT's Attentioncited by 1,003
  2. 2020A Primer in BERTology: What we know about how BERT workscited by 815
  3. 2019Designing and Interpreting Probes with Control Taskscited by 640
  4. 2020Underspecification Presents Challenges for Credibility in Modern Machine Learningcited by 496
  5. 2022Git Re-Basin: Merging Models modulo Permutation Symmetriescited by 471
  6. 2020Fine-Tuning Pretrained Language Models: Weight Initializations, Data Orders, and Early Stoppingcited by 384
  7. 2019Revealing the Dark Secrets of BERTcited by 301
  8. 2021Towards Causal Representation Learningcited by 209
  9. 2019Emerging Cross-lingual Structure in Pretrained Language Modelscited by 181
  10. 2020What Happens To BERT Embeddings During Fine-tuning?cited by 113
  11. 2021The MultiBERTs: BERT Reproductions for Robustness Analysiscited by 92
  12. 2021Similarity and Matching of Neural Network Representationscited by 78
  13. 2021Counterfactual Interventions Reveal the Causal Effect of Relative Clause Representations on Agreement Predictioncited by 59
  14. 2022Probing for the Usage of Grammatical Numbercited by 59
  15. 2021Are Larger Pretrained Language Models Uniformly Better? Comparing Performance at the Instance Levelcited by 34
  16. 2020Quantifying the Contextualization of Word Representations with Semantic Class Probingcited by 18
  17. 2019Correlations between Word Vector Setscited by 2

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2023: 1; 2024: 3; 2025: 1; 2026: 1.

6 papers.

  1. 2024Merging Text Transformer Models from Different Initializationscited by 16
  2. 2023Understanding the Inner Workings of Language Models Through Representation Dissimilaritycited by 12
  3. 2024Exploring Dark Knowledge under Various Teacher Capacities and Addressing Capacity Mismatchcited by 4
  4. 2024On Affine Homotopy between Language Encoderscited by 3
  5. 2025From Embeddings to Equations: Genetic-Programming Surrogates for Interpretable Transformer Classificationcited by 1
  6. 2026Structure Retention in Embedding Spaces as a Predictor of Benchmark Performancecited by 0