Problems and opportunities in training deep learning software systems: an analysis of variance
The DOI matches its Crossref record; 116 of the 116 references it cites were found in a registry or an index.
Abstract
Deep learning (DL) training algorithms utilize nondeterminism to improve models' accuracy and training efficiency. Hence, multiple identical training runs (e.g., identical training data, algorithm, and network) produce different models with different accuracies and training times. In addition to these algorithmic factors, DL libraries (e.g., TensorFlow and cuDNN) introduce additional variance (referred to as implementation-level variance) due to parallelism, optimization, and floating-point computation.
Overview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474 · cited 4×
“For example, the time to convergence of the 16 fixed-seed identical training runs of another popular network, ResNet56 [51], ranges”
BackgroundMethodInfluentialVery Deep Convolutional Networks for Large-Scale Image Recognition
Simonyan, K., & Zisserman, A. (2014) · arXiv · doi:10.48550/arxiv.1409.1556 · Cited by 114,148 · cited 1×
“Following common practice [14, 24, 45, 51, 62, 97, 102], we use the validation set to select the best model.”
MethodImageNet classification with deep convolutional neural networks
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017) · Communications of the ACM · doi:10.1145/3065386 · Cited by 112,155 · cited 1×
“Following common practice [14, 24, 45, 51, 62, 97, 102], we use the validation set to select the best model.”
MethodLeCun, Y., Bengio, Y., & Hinton, G. (2015) · Nature · doi:10.1038/nature14539 · Cited by 84,126
Statistical power analysis for the behavioral sciences
(1990) · Computers, Environment and Urban Systems · doi:10.1016/0198-9715(90)90050-4 · Cited by 65,269
Gradient-based learning applied to document recognition
Lecun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998) · Proceedings of the IEEE · doi:10.1109/5.726791 · Cited by 58,805 · cited 2×
“For example, our experiments show that for 16 identical training runs of a popular DL network, LeNet5 [65], the accuracy of the resulting 16 models ranges from 8.6% to 99.0% —a large accuracy difference of 90.4%.”
BackgroundMethodInfluential(2014) · Cited by 33,419
Rethinking the Inception Architecture for Computer Vision
Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.308 · Cited by 32,138
PyTorch: An Imperative Style, High-Performance Deep Learning Library
Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al. (2019) · arXiv · doi:10.48550/arxiv.1912.01703 · Cited by 31,473 · cited 4×
“DL implementation consists of 1 https://github.com/lin-tan/dl-variance high-level DL libraries (e.g., Keras), core DL libraries (e.g., Tensor-Flow and PyTorch), low-level computation libraries (e.g., cuDNN and CUDA), and hardware (e.g., GPU, CPU, and TPU).”
MethodInfluentialLearning Multiple Layers of Features from Tiny Images
Krizhevsky, A. (2024) · doi:10.57702/zp44cu3g · Cited by 23,395
Research works printed but not linked: 19
- 2019. ASE '19: Proceedings of the 34th ACM/IEEE International Conference on Automated Software Engineering.
- 2019. ASPLOS '19: Proceedings of the Twenty-Fourth International Conference on Architectural Support for Programming Languages and Operating Systems.
- 2019. CVPR'19: Proceedings of The IEEE Conference on Computer Vision and Pattern Recognition.
- 2019. ESEC/FSE '19: Proceedings of the 27th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering.
- 2019. ICCV'19: Proceedings of The IEEE International Conference on Computer Vision. IEEE Press.
- 2019. ICML'19: Proceedings of The International Conference on Machine Learning.
- 2019. ICSE '19: Proceedings of the 41st International Conference on Software Engineering. IEEE Press.
- 2019. MLSys'19: Proceedings of Machine Learning and Systems.
- 2019. NIPS'19: Proceedings of the 33rd Conference on Neural Information Processing Systems.
- 2019. SOSP '19: Proceedings of the 27th ACM Symposium on Operating Systems Principles.
- 2020. ICLR'20: Proceedings of The International Conference on Learning Representations.
- 2020. Reproducibility Challenge @ NeurIPS 2019. https://reproducibilitychallenge.github.io/neurips2019/
- Charu C Aggarwal. 2018. Neural networks and deep learning. In Springer.
- Cody Coleman, Deepak Narayanan, Daniel Kang, Tian Zhao, Jian Zhang, Luigi Nardi, Peter Bailis, Kunle Olukotun, Chris Ré, and Matei Zaharia. 2017. Dawnbench: An end-to-end deep learning benchmark and competition. In Training.
- Standard Performance Evaluation Corporation. 2006. SPEC CPU2006 Benchmarks. http://www.spec.org/cpu2006/.
- Heming Cui, Jiri Simsa, Yi-Hong Lin, Hao Li, Ben Blum, Xinan Xu, Junfeng Yang, Garth A. Gibson, and Randal E. Bryant. 2013. Parrot: a Practical Runtime for Deterministic, Stable, and Reliable Threads. In SOSP.
- Jeff Haochen and Suvrit Sra. 2019. Random Shuffling Beats SGD after Finite Epochs. In ICML.
- Hadi Jooybar, Wilson WL Fung, Mike O'Connor, Joseph Devietti, and Tor M Aamodt. 2013. GPUDet: a deterministic GPU architecture. In ASPLOS.
- Shlomo Sawilowsky, Jack Sawilowsky, and Robert J. Grissom. 2011. Effect Size.
Cited by
Machine learning for medical imaging: methodological failures and recommendations for the future
Varoquaux, G., & Cheplygina, V. (2022) · npj Digital Medicine · doi:10.1038/s41746-022-00592-y · Cited by 656
An Empirical Study of the Non-Determinism of ChatGPT in Code Generation
Ouyang, S., Zhang, J. M., Harman, M., & Wang, M. (2025) · ACM Transactions on Software Engineering and Methodology · doi:10.1145/3697010 · Cited by 258
Giray, G. (2021) · Journal of Systems and Software · doi:10.1016/j.jss.2021.111031 · Cited by 187
An extensive study on pre-trained models for program understanding and generation
Zeng, Z., Tan, H., Zhang, H., Li, J., Zhang, Y., & Zhang, L. (2022) · Proceedings of the 31st ACM SIGSOFT International Symposium on Software Testing and Analysis · doi:10.1145/3533767.3534390 · Cited by 153
“Note that considering the model randomness, we run all our experiments for 5 runs as recommended by prior work [52] and present the average results, while all our studied pre-trained models do not clearly indicate their number of runs in their original papers.”
MethodYan, J., Cai, J., Zhang, B., Wang, Y., Wong, D. F., & Siu, S. W. I. (2022) · Antibiotics · doi:10.3390/antibiotics11101451 · Cited by 146
Metadata
- Authors
- Pham, H. V., Qian, S., Wang, J., Lutellier, T., Rosenthal, J., Tan, L., Yu, Y., & Nagappan, N.
- Affiliations
- University of Waterloo (CA)
- Purdue University West Lafayette (US)
- Microsoft (United States) (US)
- Venue
- Proceedings of the 35th IEEE/ACM International Conference on Automated Software Engineering
- Publisher
- ACM
- Published
- 2020-12-21
- Pages
- 771-783
- OpenAlex
- W3122184684
- Cited by
- 100
- References
- 114
- Citations by year
- 2026: 13 · 2025: 11 · 2024: 16 · 2023: 29 · 2022: 19 · 2021: 11
- FWCI
- 9.79
- Type
- conference paper
- Language
- en
- Open access
- closed
- Locations
- IEEE/ACM International Conference on Automated Software Engineering (ASE) (publishedVersion)
- Funders
- National science foundation
- Facebook Research Award
- Topics
- Physical Sciences › Computer Science › Software › Software Testing and Debugging Techniques
- Physical Sciences › Computer Science › Hardware and Architecture › Parallel Computing and Optimization Techniques
- Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning and Data Classification
- Keywords
- deep learning training, non-determinism, model reproducibility, parallelism, floating-point arithmetic, training efficiency, TensorFlow, cuDNN
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: Problems and opportunities in training deep learning software systems: an analysis of variance. https://app.cochat.ai/papers/10.1145/3324884.3416545
- The paper
- Pham, H. V., Qian, S., Wang, J., Lutellier, T., Rosenthal, J., Tan, L., Yu, Y., & Nagappan, N. (2020). Problems and opportunities in training deep learning software systems: an analysis of variance. Proceedings of the 35th IEEE/ACM International Conference on Automated Software Engineering. https://doi.org/10.1145/3324884.3416545
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.