Elamur Research⌕
☰
home / papers / 2020 / Estimating Training Data Influence by Tracing Gradient Descent

Estimating Training Data Influence by Tracing Gradient Descent

Paper · 2020 · Garima Pruthi, Frederick Liu, Mukund Sundararajan et al.

Summary

The paper asks which training examples are responsible for a model's prediction on a given input. The authors propose TracIn. It attributes to each training example the drop in the test example's loss that occurred during the optimization steps where that example was used. The exact version would require knowing the test point while training, so they derive a practical variant. A first-order expansion turns each step's contribution into a learning-rate-weighted dot product between the training-example gradient and the test-example gradient. Saved checkpoints stand in for the full training trajectory (TracInCP), and gradients can be restricted to selected layers to reduce cost. The method only needs gradients, checkpoints and a loss. It therefore applies to any model trained with SGD-like optimizers.

The authors compare it with influence functions and representer point selection, using self-influence to surface mislabelled training data. They also demonstrate it on regression, text and image tasks. The result is an easy-to-implement tool for auditing and cleaning training sets.

Key points

  • Two variants:
    • TracIn (first-order, per-step) sums η_t·∇ℓ(w_t, z')·∇ℓ(w_t, z) over the steps that used z, with averaging for minibatches.
    • TracInCP replaces per-step parameters with k stored checkpoints.
  • Evaluation: training points are ranked by self-influence, and the metric is the share of mislabelled examples recovered in top prefixes. Baselines are influence functions and representer points.
  • DBPedia (text): 560K training and 70K test examples across 14 classes, using a SWEM model with a sentencepiece tokenizer. The model reached about 95.5% accuracy. Influence was computed from 6 checkpoints on the last fully connected layer.
  • ImageNet: ResNet-50 at 73% top-1, trained on 1.28M images in 1000 classes. TracInCP used checkpoints 30/60/90, with gradients projected to 1472 dimensions via a rank-1 trick for fully connected layers.
  • Housing-price regression: a model with 168K parameters, 3 hidden layers and test explained variance of 0.70. High self-influence points came from densely populated areas, which the authors link to memorization.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 works

Cite

Cite as Garima Pruthi, Frederick Liu, Mukund Sundararajan and Satyen Kale. 2020. Estimating Training Data Influence by Tracing Gradient Descent. arXiv:2002.08484.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2002.08484 · arXiv 2002.08484

@misc{pruthi2020estimating,
  title = {{Estimating Training Data Influence by Tracing Gradient Descent}},
  author = {Garima Pruthi, Frederick Liu, Mukund Sundararajan and Satyen Kale},
  year = {2020},
  doi = {10.48550/arxiv.2002.08484},
  eprint = {2002.08484},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2002.08484}
}

References

28 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. TracIn Code. https://github.com/frederick0329/TracIn. ↗
  2. Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 1885–1894, 2017.
  3. Chih-Kuan Yeh, Joon Kim, Ian En-Hsu Yen, and Pradeep K Ravikumar. Representer point selection for explaining deep neural networks. In Advances in Neural Information Processing Systems, pages 9291–9301, 2018. find
  4. Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3319–3328, 2017. find
  5. Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Advances in neural information processing systems, pages 4765–4774, 2017.
  6. Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. " Why should i trust you?" Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135–1144, 2016.
  7. Art B Owen and Clémentine Prieur. On shapley value for measuring importance of dependent inputs. SIAM/ASA Journal on Uncertainty Quantification, 5(1):986–1002, 2017. find
  8. Art B Owen. Sobol’indices and shapley value. SIAM/ASA Journal on Uncertainty Quantification, 2(1):245–251, 2014. find
  9. Ruoxi Jia, David Dao, Boxin Wang, Frances Ann Hubis, Nick Hynes, Nezihe Merve Gürel, Bo Li, Ce Zhang, Dawn Song, and Costas J Spanos. Towards efficient data valuation based on the shapley value. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 1167–1176, 2019. find
  10. Amirata Ghorbani and James Zou. Data shapley: Equitable valuation of data for machine learning. In International Conference on Machine Learning, pages 2242–2251, 2019. find
  11. Satoshi Hara, Atsushi Nitanda, and Takanori Maehara. Data cleansing for models trained with sgd. In Advances in Neural Information Processing Systems, pages 4215–4224, 2019. find
  12. Rico Sennrich, Barry Haddow, and Alexandra Birch. Edinburgh neural machine translation systems for WMT 16. In Proceedings of the First Conference on Machine Translation, pages 371–376, 2016.
  13. Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In 5th International Conference on Learning Representations, 2017. find
  14. Hao-Ran Wei, Shujian Huang, Ran Wang, Xin-Yu Dai, and Jiajun Chen. Online distilling from checkpoints for neural machine translation. In Jill Burstein, Christy Doran, and Thamar Solorio, editors, Proceedings of the 17th Conference of the North American Chapter of the Association for Computational Li… find
  15. Yanyao Shen and Sujay Sanghavi. Learning with bad training data via iterative trimmed loss minimization. In ICML, 2019. find
  16. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016.
  17. Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, Toronto, Canada, 2009. find
  18. R Kelley Pace and Ronald Barry. Sparse spatial autoregressions. Statistics & Probability Letters, 33(3):291–297, 1997. find
  19. Xiang Zhang, Junbo Zhao, and Yann LeCun. Character-level convolutional networks for text classification. In Advances in neural information processing systems, pages 649–657, 2015.
  20. Dinghan Shen, Guoyin Wang, Wenlin Wang, Martin Renqiang Min, Qinliang Su, Yizhe Zhang, Chunyuan Li, Ricardo Henao, and Lawrence Carin. Baseline needs more love: On simple word-embedding-based models and associated pooling mechanisms. In Proceedings of the 56th Annual Meeting of the Association for C…
  21. Taku Kudo and John Richardson. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66–71, 2018.
  22. Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. find
  23. Xiaochuang Han and Yulia Tsvetkov. Fortifying toxic speech detectors against veiled toxicity, 2020. find
  24. R Dennis Cook and Sanford Weisberg. Residuals and influence in regression. New York: Chapman and Hall, 1982. find
  25. Barak A. Pearlmutter. Fast exact multiplication by the hessian. Neural Computation, 6:147–160, 1994. find
  26. Bernhard Schölkopf, Ralf Herbrich, and Alex J Smola. A generalized representer theorem. In International conference on computational learning theory, pages 416–426. Springer, 2001. find
  27. Erik Bernhardsson. Annoy: Approximate Nearest Neighbors in C++/Python, 2018. Python package version 1.13.0. find
  28. David P Woodruff et al. Sketching as a tool for numerical linear algebra. Foundations and Trends® in Theoretical Computer Science, 10(1–2):1–157, 2014. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/pruthi-2020-estimating-training-data-influence-tracing/