2021

Editing a classifier by rewriting its prediction rules

Santurkar, Shibani, Tsipras, Dimitris, Elango, Mahalaxmi et al.

Understand

We present a methodology for modifying the behavior of a classifier by directly rewriting its prediction rules.

Our approach requires virtually no additional data collection and can be applied to a variety of settings, including adapting a model to new environments, and modifying it to ignore spurious features.
Our code is available at https://github.com/MadryLab/EditingClassifiers .

Built on

“Dataset issues in object recognition”
Jean Ponce et al · 2006
Earlier work this paper cites.
“Analysis of representations for domain adaptation”
Shai Ben-David, John Blitzer, Koby Crammer and Fernando Pereira · 2007
Earlier work this paper cites.
“Local features and kernels for classification of texture and object categories: A comprehensive study”
Jianguo Zhang, Marcin Marszałek, Svetlana Lazebnik and Cordelia Schmid · 2007
Earlier work this paper cites.
“Imagenet: A large-scale hierarchical image database”
Jia Deng et al · 2009
Earlier work this paper cites.
“Visualizing higher-layer features of a deep network”
Dumitru Erhan, Yoshua Bengio, Aaron Courville and Pascal Vincent · 2009
Earlier work this paper cites.
“Learning to detect unseen object classes by between-class attribute transfer”
Christoph Lampert, Hannes Nickisch and Stefan Harmeling · 2009
Earlier work this paper cites.
“Causal inference”
Judea Pearl · 2010
Earlier work this paper cites.
“Adapting visual category models to new domains”
Kate Saenko, Brian Kulis, Mario Fritz and Trevor Darrell · 2010
Earlier work this paper cites.
“Unbiased look at dataset bias”
Antonio Torralba and Alexei Efros · 2011
Earlier work this paper cites.
“Undoing the damage of dataset bias”
Aditya Khosla et al · 2012
Earlier work this paper cites.
“Decaf: A deep convolutional activation feature for generic visual recognition”
Jeff Donahue et al · 2014
Earlier work this paper cites.
“Microsoft coco: Common objects in context”
Tsung-Yi Lin et al · 2014
Earlier work this paper cites.
“CNN features off-the-shelf: an astounding baseline for recognition”
Ali Sharif, Hossein Azizpour, Josephine Sullivan and Stefan Carlsson · 2014
Earlier work this paper cites.
“Intriguing properties of neural networks”
Christian Szegedy et al · 2014
Earlier work this paper cites.
“A testbed for cross-dataset analysis”
Tatiana Tommasi and Tinne Tuytelaars · 2014
Earlier work this paper cites.
“Visualizing and understanding convolutional networks”, 2014
Matthew Zeiler and Rob Fergus · 2014
Earlier work this paper cites.
“Manitest: Are classifiers really invariant?”
Alhussein Fawzi and Pascal Frossard · 2015
Earlier work this paper cites.
“Unsupervised domain adaptation by backpropagation”
Yaroslav Ganin and Victor Lempitsky · 2015
Earlier work this paper cites.
“Deep Residual Learning for Image Recognition”
Kaiming He, Xiangyu Zhang, Shaoqing Ren and Jian Sun · 2015
Earlier work this paper cites.
“ImageNet Large Scale Visual Recognition Challenge”
Olga Russakovsky et al · 2015
Earlier work this paper cites.
“Very Deep Convolutional Networks for Large-Scale Image Recognition”
Karen Simonyan and Andrew Zisserman · 2015
Earlier work this paper cites.
“Optimal transport for domain adaptation”
Nicolas Courty, Rémi Flamary, Devis Tuia and Alain Rakotomamonjy · 2016
Earlier work this paper cites.
“Robustness of classifiers: from adversarial to random noise”
Alhussein Fawzi, Seyed-Mohsen Moosavi-Dezfooli and Pascal Frossard · 2016
Earlier work this paper cites.
“Image style transfer using convolutional neural networks”
Leon Gatys, Alexander Ecker and Matthias Bethge · 2016
Earlier work this paper cites.
“Domain adaptation with conditional transferable components”
Mingming Gong et al · 2016
Earlier work this paper cites.
“Deep Residual Learning for Image Recognition”
Kaiming He, Xiangyu Zhang, Shaoqing Ren and Jian Sun · 2016
Earlier work this paper cites.
“Revisiting batch normalization for practical domain adaptation”
Original
Yanghao Li et al · 2016
Earlier work this paper cites.
““Why Should I Trust You?”: Explaining the Predictions of Any Classifier”
Marco Ribeiro, Sameer Singh and Carlos Guestrin · 2016
Earlier work this paper cites.
“Network dissection: Quantifying interpretability of deep visual representations”
David Bau et al · 2017
Earlier work this paper cites.
“Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs”
Liang-Chieh Chen et al · 2017
Earlier work this paper cites.

Similar

“Exploring the structure of a real-time, arbitrary neural artistic stylization network”
Original
Golnaz Ghiasi et al · 2017
Cited alongside, same era.
“Conditional variance penalties and domain shift robustness”
Original
Christina Heinze-Deml and Nicolai Meinshausen · 2017
Cited alongside, same era.
“Few-Shot Adversarial Domain Adaptation”
Saeid Motiian, Quinn Jones, Seyed Iranmanesh and Gianfranco Doretto · 2017
Cited alongside, same era.
“Feature Visualization”
Chris Olah, Alexander Mordvintsev and Ludwig Schubert · 2017
Cited alongside, same era.
“Places: A 10 million image database for scene recognition”
Bolei Zhou et al · 2017
Cited alongside, same era.
“Recognition in terra incognita”
Sara Beery, Grant Van and Pietro Perona · 2018
Cited alongside, same era.
“Detectron”, https://github.com/facebookresearch/detectron , 2018
Ross Girshick et al · 2018
Cited alongside, same era.
“Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav)”
Been Kim et al · 2018
Cited alongside, same era.
“Towards deep learning models resistant to adversarial attacks”
Aleksander Madry et al · 2018
Cited alongside, same era.
“The Building Blocks of Interpretability”
Chris Olah et al · 2018
Cited alongside, same era.
“The Elephant in the Room”
Original
Amir Rosenfeld, Richard Zemel and John. Tsotsos · 2018
Cited alongside, same era.
“Convnets and imagenet beyond accuracy: Understanding mistakes and uncovering biases”
Pierre Stock and Moustapha Cisse · 2018
Cited alongside, same era.

Then

“Improving robustness without sacrificing accuracy with patch gaussian augmentation”
Original
Raphael Lopes et al · 2019
Later among the works it cites.
“Distributionally Robust Language Modeling”
Yonatan Oren, Shiori Sagawa, Tatsunori Hashimoto and Percy Liang · 2019
Later among the works it cites.
“Do ImageNet Classifiers Generalize to ImageNet?”
Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt and Vaishaal Shankar · 2019
Later among the works it cites.
“Not Using the Car to See the Sidewalk–Quantifying and Controlling the Effects of Context in Classification and Segmentation”
Rakshith Shetty, Bernt Schiele and Mario Fritz · 2019
Later among the works it cites.
“A fourier perspective on model robustness in computer vision”
Dong Yin et al · 2019
Later among the works it cites.
“Towards causal vqa: Revealing and reducing spurious correlations by invariant and covariant semantic editing”
Vedika Agarwal, Rakshith Shetty and Mario Fritz · 2020
Later among the works it cites.
“Are we done with ImageNet?”
Original
Lucas Beyer et al · 2020
Later among the works it cites.
“Rewriting a deep generative model”
David Bau et al · 2020
Later among the works it cites.
“Debiasing Skin Lesion Datasets and Models? Not So Fast”
Alceu Bissoto, Eduardo Valle and Sandra Avila · 2020
Later among the works it cites.
“Understanding the role of individual units in a deep neural network”
David Bau et al · 2020
Later among the works it cites.
“Concept whitening for interpretable image recognition”
Zhi Chen, Yijie Bei and Cynthia Rudin · 2020
Later among the works it cites.
“Shortcut learning in deep neural networks”
Robert Geirhos et al · 2020
Later among the works it cites.
“GANSpace: Discovering Interpretable GAN Controls”
Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen and Sylvain Paris · 2020
Later among the works it cites.
“Understanding Self-Training for Gradual Domain Adaptation”
Ananya Kumar, Tengyu Ma and Percy Liang · 2020
Later among the works it cites.
“Concept bottleneck models”
Pang Koh et al · 2020
Later among the works it cites.
“WILDS: A Benchmark of in-the-Wild Distribution Shifts”
Original
Pang Koh et al · 2020
Later among the works it cites.
“Interpreting the latent space of gans for semantic face editing”
Yujun Shen, Jinjin Gu, Xiaoou Tang and Bolei Zhou · 2020
Later among the works it cites.
“Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization”
Shiori Sagawa, Pang Koh, Tatsunori. Hashimoto and Percy Liang · 2020
Later among the works it cites.
“From ImageNet to Image Classification: Contextualizing Progress on Benchmarks”
Dimitris Tsipras et al · 2020
Later among the works it cites.
“StyleSpace Analysis: Disentangled Controls for StyleGAN Image Generation”
Zongze Wu, Dani Lischinski and Eli Shechtman · 2020
Later among the works it cites.
“Noise or signal: The role of image backgrounds in object recognition”
Original
Kai Xiao, Logan Engstrom, Andrew Ilyas and Aleksander Madry · 2020
Later among the works it cites.
“Limitations of Post-Hoc Feature Alignment for Robustness”
Collin Burns and Jacob Steinhardt · 2021
Closest in time.
“Editing Factual Knowledge in Language Models”
Original
Nicola De, Wilker Aziz and Ivan Titov · 2021
Closest in time.
“Knowledge neurons in pretrained transformers”
Original
Damai Dai et al · 2021
Closest in time.
“Multimodal neurons in artificial neural networks”
Gabriel Goh et al · 2021
Closest in time.
“Fast Model Editing at Scale”
Original
Eric Mitchell et al · 2021
Closest in time.
“Learning transferable visual models from natural language supervision”
Original
Alec Radford et al · 2021
Closest in time.
“Breeds: Benchmarks for subpopulation shift”
Shibani Santurkar, Dimitris Tsipras and Aleksander Madry · 2021
Closest in time.
“Leveraging Sparse Linear Layers for Debuggable Deep Networks”
Eric Wong, Shibani Santurkar and Aleksander Madry · 2021
Closest in time.
“How Does Mixup Help With Robustness and Generalization?”
Linjun Zhang et al · 2021
Closest in time.

Beyond the bibliography

alphaXiv searches the wider corpus for related work and actual follow-ups.

Open on alphaXiv

alphaXiv is searching for related work…

Reading the bibliography…