Data-Centric Debugging: mitigating model failures via targeted data collection
Not on the map. Why?
Abstract (arXiv)
Deep neural networks can be unreliable in the real world when the training set does not adequately cover all the settings where they are deployed. Focusing on image classification, we consider the setting where we have an error distribution E representing a deployment scenario where the model fails. We have access to a small set of samples Esample from E and it can be expensive to obtain additional samples. In the traditional model development framework, mitigating failures of the model in E can be challenging and is often done in an ad hoc manner. In this paper, we propose a general methodology for model debugging that can systemically improve model performance on E while maintaining its performance on the original test set. Our key assumption is that we have access to a large pool of weakly (noisily) labeled data F. However, naively adding F to the training would hurt model performance due to the large extent of label noise. Our Data-Centric Debugging (DCD) framework carefully creates a debug-train set by selecting images from F that are perceptually similar to the images in Esample. To do this, we use the ℓ2 distance in the feature space (penultimate layer activations) of various models including ResNet, Robust ResNet and DINO where we observe DINO ViTs are significantly better at discovering similar images compared to Resnets. Compared to LPIPS, we find that our method reduces compute and storage requirements by 99.58%. Compared to the baselines that maintain model performance on the test set, we achieve significantly (+9.45%) improved results on the debug-heldout sets.
Key facts. Cited by 3 papers in the map (data through 13 Oct 2026). Builds on 54 of its 72 references that are in the map. First posted on arXiv 17 Nov 2022.
Builds on
54 of 72 references.
References · 54
- 2015Deep Residual Learning for Image Recognitioncited by 51,950
- 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,304
- 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,588
- 2020Language Models are Few-Shot Learnerscited by 32,320
- 2020An Image is Worth 16x16 Words: Transformers for Image Recognition at Scalecited by 30,043
- 2020A Simple Framework for Contrastive Learning of Visual Representationscited by 12,009
- 2018The Unreasonable Effectiveness of Deep Features as a Perceptual Metriccited by 11,082
- 2019Momentum Contrast for Unsupervised Visual Representation Learningcited by 7,492
- 2021Masked Autoencoders Are Scalable Vision Learnerscited by 7,364
- 2021Emerging Properties in Self-Supervised Vision Transformerscited by 6,915
- 2016Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localizationcited by 6,713
- 2016"Why Should I Trust You?": Explaining the Predictions of Any Classifiercited by 5,055
38 more
- 2017Axiomatic Attribution for Deep Networkscited by 3,368
- 2019Benchmarking Neural Network Robustness to Common Corruptions and Perturbationscited by 2,862
- 2015Learning Deep Features for Discriminative Localizationcited by 2,434
- 2013Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Mapscited by 2,360
- 2013Visualizing and Understanding Convolutional Networkscited by 2,282
- 2020The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalizationcited by 1,847
- 2017Understanding Black-box Predictions via Influence Functionscited by 1,729
- 2019Do ImageNet Classifiers Generalize to ImageNet?cited by 1,510
- 2019Natural Adversarial Examplescited by 1,405
- 2018Sanity Checks for Saliency Mapscited by 1,152
- 2017SmoothGrad: removing noise by adding noisecited by 1,103
- 2014Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Imagescited by 774
- 2021Editing Factual Knowledge in Language Modelscited by 551
- 2021Fast Model Editing at Scalecited by 493
- 2014Understanding Deep Image Representations by Inverting Themcited by 400
- 2018Confounding variables can degrade generalization performance of radiological deep learning modelscited by 375
- 2015Understanding Neural Networks Through Deep Visualizationcited by 333
- 2019Counterfactual Visual Explanationscited by 313
- 2019On the (In)fidelity and Sensitivity for Explanationscited by 310
- 2020Understanding the Role of Individual Units in a Deep Neural Networkcited by 255
- 2021Combined Scaling for Zero-shot Transfer Learningcited by 182
- 2015Inverting Visual Representations with Convolutional Networkscited by 168
- 2018Explaining Image Classifiers by Counterfactual Generationcited by 144
- 2021Salient ImageNet: How to discover spurious features in Deep Learning?cited by 114
- 2016Multifaceted Feature Visualization: Uncovering the Different Types of Features Learned By Each Neuron in Deep Neural Networkscited by 106
- 2017Interpreting Deep Visual Representations via Network Dissectioncited by 103
- 2020Benchmarking Deep Learning Interpretability in Time Series Predictionscited by 102
- 2022Quality Not Quantity: On the Interaction between Dataset Design and Robustness of CLIPcited by 97
- 2015Visualizing Deep Convolutional Neural Networks Using Natural Pre-Imagescited by 93
- 2021Leveraging Sparse Linear Layers for Debuggable Deep Networkscited by 82
- 2018Automated Data Slicing for Model Validation:A Big data - AI Integration Approachcited by 70
- 2021Editing a classifier by rewriting its prediction rulescited by 65
- 2020Understanding Failures of Deep Networks via Robust Feature Extractioncited by 56
- 2018Manifold: A Model-Agnostic Framework for Interpretation and Diagnosis of Machine Learning Modelscited by 50
- 2018Towards Accountable AI: Hybrid Human-Machine Analyses for Characterizing System Failurecited by 49
- 2020Generative causal explanations of black-box classifierscited by 41
- 2022Missingness Bias in Model Debuggingcited by 38
- 20213DB: A Framework for Debugging Computer Vision Modelscited by 31
50 most cited of 54 All 54 (JSON)
Built on by
Citing papers per year: 2023: 2; 2024: 1.
3 papers.