mlsanityData to 14 Oct 2026

Deconstructing Distributions: A Pointwise Framework of Learning

arXiv 2202.09931 ↗First posted 20 Feb 2022cs.LG cs.AI cs.CVcited by 19

Not on the map. Why?

Abstract (arXiv)

In machine learning, we traditionally evaluate the performance of a single model, averaged over a collection of test inputs. In this work, we propose a new approach: we measure the performance of a collection of models when evaluated on a textitsingle input point. Specifically, we study a point's textitprofile: the relationship between models' average performance on the test distribution and their pointwise performance on this individual point. We find that profiles can yield new insights into the structure of both models and data -- in and out-of-distribution. For example, we empirically show that real data distributions consist of points with qualitatively different profiles. On one hand, there are "compatible" points with strong correlation between the pointwise and average performance. On the other hand, there are points with weak and even textitnegative correlation: cases where improving overall model accuracy actually textithurts performance on these inputs. We prove that these experimental observations are inconsistent with the predictions of several simplified models of learning proposed in prior work. As an application, we use profiles to construct a dataset we call CIFAR-10-NEG: a subset of CINIC-10 such that for standard models, accuracy on CIFAR-10-NEG is textitnegatively correlated with accuracy on CIFAR-10 test. This illustrates, for the first time, an OOD dataset that completely inverts "accuracy-on-the-line" (Miller, Taori, Raghunathan, Sagawa, Koh, Shankar, Liang, Carmon, and Schmidt 2021)

Key facts. Cited by 19 papers in the map (data through 14 Oct 2026). Builds on 57 of its 92 references that are in the map. First posted on arXiv 20 Feb 2022.

Builds on

57 of 92 references.

References · 57

  1. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,399
  2. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,744
  3. 2020Language Models are Few-Shot Learnerscited by 32,419
  4. 2020Measuring Massive Multitask Language Understandingcited by 7,963
  5. 2020Scaling Laws for Neural Language Modelscited by 6,586
  6. 2021Measuring Mathematical Problem Solving With the MATH Datasetcited by 5,371
  7. 2016Simple and Scalable Predictive Uncertainty Estimation using Deep Ensemblescited by 3,926
  8. 2016A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networkscited by 2,385
  9. 2020The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalizationcited by 1,850
  10. 2019Do ImageNet Classifiers Generalize to ImageNet?cited by 1,515
  11. 2019Natural Adversarial Examplescited by 1,408
  12. 2017Enhancing The Reliability of Out-of-distribution Image Detection in Neural Networkscited by 1,328
38 more
  1. 2019Can You Trust Your Model's Uncertainty? Evaluating Predictive Uncertainty Under Dataset Shiftcited by 1,287
  2. 2019Learning Robust Global Representations by Penalizing Local Predictive Powercited by 1,013
  3. 2021Measuring Coding Challenge Competence With APPScited by 902
  4. 2017Selective Classification for Deep Neural Networkscited by 852
  5. 2018An Empirical Study of Example Forgetting during Deep Neural Network Learningcited by 678
  6. 2019Deep Double Descent: Where Bigger Models and More Data Hurtcited by 672
  7. 2022Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasetscited by 629
  8. 2016Training Region-based Object Detectors with Online Hard Example Miningcited by 579
  9. 2020Scaling Laws for Autoregressive Generative Modelingcited by 522
  10. 2021Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarkscited by 395
  11. 2017A Downsampled Variant of ImageNet as an Alternative to the CIFAR datasetscited by 377
  12. 2018Detecting and Correcting for Label Shift with Black Box Predictorscited by 366
  13. 2021Explaining Neural Scaling Lawscited by 363
  14. 2020Calibration of Pre-trained Transformerscited by 315
  15. 2020Pretrained Transformers Improve Out-of-Distribution Robustnesscited by 294
  16. 2019On The Power of Curriculum Learning in Training Deep Networkscited by 286
  17. 2020Identifying Mislabeled Data using the Area Under the Margin Rankingcited by 261
  18. 2021Accuracy on the Line: On the Strong Correlation Between Out-of-Distribution and In-Distribution Generalizationcited by 249
  19. 2019A Constructive Prediction of the Generalization Error Across Scalescited by 210
  20. 2018To Trust Or Not To Trust A Classifiercited by 207
  21. 2019SGD on Neural Networks Learns Functions of Increasing Complexitycited by 204
  22. 2019What Do Compressed Deep Neural Networks Forget?cited by 171
  23. 2022Leveraging Unlabeled Data to Predict Out-of-Distribution Performancecited by 145
  24. 2021Deep Learning Through the Lens of Example Difficultycited by 141
  25. 2020A Unified View of Label Shift Estimationcited by 119
  26. 2021Predicting with Confidence on Unseen Distributionscited by 116
  27. 2020Optimal Regularization Can Mitigate Double Descentcited by 109
  28. 2020Hybrid Models for Open Set Recognitioncited by 109
  29. 2020Are Labels Always Necessary for Classifier Accuracy Evaluation?cited by 104
  30. 2020Estimating Example Difficulty Using Variance of Gradientscited by 97
  31. 2022Limitations of Neural Collapse for Understanding Generalization in Deep Learningcited by 63
  32. 2021No One Representation to Rule Them All: Overlapping Features of Training Methodscited by 58
  33. 2020A Neural Scaling Law from the Dimension of the Data Manifoldcited by 52
  34. 2020Distributional Generalization: A New Kind of Generalizationcited by 42
  35. 2019Model Similarity Mitigates Test Set Overusecited by 38
  36. 2021Are Larger Pretrained Language Models Uniformly Better? Comparing Performance at the Instance Levelcited by 34
  37. 2021The Three Stages of Learning Dynamics in High-Dimensional Kernel Methodscited by 22
  38. 2020Learning Rate Annealing Can Provably Help Generalization, Even for Convex Problemscited by 11

50 most cited of 57 All 57 (JSON)

Built on by

Cited by year

Citing papers per year: 2022: 7; 2023: 5; 2024: 3; 2025: 4.

19 papers.

Sort:

By area:

  1. 2022Agreement-on-the-Line: Predicting the Performance of Neural Networks under Distribution Shiftcited by 94
  2. 2022ID and OOD Performance Are Sometimes Inversely Correlated on Real-world Datasetscited by 44
  3. 2023T-MARS: Improving Visual Representations by Circumventing Text Feature Learningcited by 43
  4. 2022Characterizing Datapoints via Second-Split Forgettingcited by 37
  5. 2022Metadata Archaeology: Unearthing Data Subsets by Leveraging Training Dynamicscited by 26
  6. 2023Tools for Verifying Neural Models' Training Datacited by 22
  7. 2023Unlocking Accuracy and Fairness in Differentially Private Image Classificationcited by 16
  8. 2023Understanding the Detrimental Class-level Effects of Data Augmentationcited by 13
  9. 2022The Value of Out-of-Distribution Datacited by 12
  10. 2024Efficient Lifelong Model Evaluation in an Era of Rapid Progresscited by 10
  11. 2022A picture of the space of typical learnable taskscited by 7
  12. 2023Accuracy on the Curve: On the Nonlinear Correlation of ML Performance Between Data Subpopulationscited by 7
7 more
  1. 2025Understanding the Mechanisms of Fast Hyperparameter Transfercited by 7
  2. 2022Rethinking Streaming Machine Learning Evaluationcited by 5
  3. 2024Reconsidering Sentence-Level Sign Language Translationcited by 4
  4. 2025The Impact of Coreset Selection on Spurious Correlations and Group Robustnesscited by 4
  5. 2025Performance Heterogeneity in Graph Neural Networks: Lessons for Architecture Design and Preprocessingcited by 2
  6. 2024Protecting against simultaneous data poisoning attackscited by 1
  7. 2025RedunCut: Measurement-Driven Sampling and Accuracy Performance Modeling for Low-Cost Live Video Analyticscited by 0