Finding and removing Clever Hans: Using explanation methods to debug and improve deep models
Open accessThe DOI matches its Crossref record; 40 of the 40 references it cites were found in a registry or an index.
Abstract
Contemporary learning models for computer vision are typically trained on very large (benchmark) datasets with millions of samples. These may, however, contain biases, artifacts, or errors that have gone unnoticed and are exploitable by the model. In the worst case, the trained model does not learn a valid and generalizable strategy to solve the problem it was trained for, and becomes a “Clever Hans” predictor that bases its decisions on spurious correlations in the training data , potentially yielding an unrepresentative or unfair, and possibly even hazardous predictor. In this paper, we contribute by providing a comprehensive analysis framework based on a scalable statistical analysis of attributions from explanation methods for large data corpora . Based on a recent technique — Spectral Relevance Analysis — we propose the following technical contributions and resulting findings: (a) a scalable quantification of artifactual and poisoned classes where the machine learning models under study exhibit Clever Hans behavior, (b) several approaches we collectively denote as Class Artifact Compensation, which are able to effectively and significantly reduce a model’s Clever Hans behavior, i.e., we are able to un-Hans models trained on (poisoned) datasets, such as the popular ImageNet data corpus. We demonstrate that Class Artifact Compensation, defined in a simple theoretical framework, may be implemented as part of a neural network’s training or fine-tuning process, or in a post-hoc manner by injecting additional layers, preventing any further propagation of undesired Clever Hans features, into the network architecture . Using our proposed methods, we provide qualitative and quantitative analyses of the biases and artifacts in, e.g., the ImageNet dataset, the Adience benchmark dataset of unfiltered faces, and the ISIC 2019 skin lesion analysis dataset. We demonstrate that these insights can give rise to improved, more representative, and fairer models operating on implicitly cleaned data corpora.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
ImageNet 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
LeCun, Y., Bengio, Y., & Hinton, G. (2015) · Nature · doi:10.1038/nature14539 · Cited by 84,126
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
Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization
Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D. (2017) · 2017 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2017.74 · Cited by 24,513
A Unified Approach to Interpreting Model Predictions
Lundberg, S., & Lee, S.-I. (2017) · arXiv · doi:10.48550/arxiv.1705.07874 · Cited by 23,707
Least squares quantization in PCM
Lloyd, S. (1982) · IEEE Transactions on Information Theory · doi:10.1109/tit.1982.1056489 · Cited by 15,836
Mastering the game of Go with deep neural networks and tree search
Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., van den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., et al. (2016) · Nature · doi:10.1038/nature16961 · Cited by 15,612
From local explanations to global understanding with explainable AI for trees
Lundberg, S. M., Erion, G., Chen, H., DeGrave, A., Prutkin, J. M., Nair, B., Katz, R., Himmelfarb, J., Bansal, N., & Lee, S.-I. (2020) · Nature Machine Intelligence · doi:10.1038/s42256-019-0138-9 · Cited by 10,551 · cited 1×
“Shapley value based approaches [57, 58] leverage tools from game theory in order to estimate the importance of features to a decision of a model.”
MethodGene selection for cancer classification using support vector machines
Guyon, I., Weston, J., Barnhill, S., & Vapnik, V. (2002) · Machine Learning · doi:10.1023/a:1012487302797 · Cited by 9,954
Towards Evaluating the Robustness of Neural Networks
Carlini, N., & Wagner, D. (2017) · 2017 IEEE Symposium on Security and Privacy (SP) · doi:10.1109/sp.2017.49 · Cited by 8,859
UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction
McInnes, L., Healy, J., & Melville, J. (2018) · arXiv · doi:10.48550/arxiv.1802.03426 · Cited by 7,101
Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps
Simonyan, K., Vedaldi, A., & Zisserman, A. (2013) · arXiv · doi:10.48550/arxiv.1312.6034 · Cited by 5,370 · cited 1×
“, with (modified) backpropagation approaches, such as sensitivity analysis [16, 54], Layer-wise Relevance Propagation (LRP) [17], Deep Taylor Decomposition [55], Grad-CAM [19], Integrated Gradients [20], SmoothGrad [56], and DeepLIFT [21], which require access to the internal parameters of DNN models.”
MethodOn Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation
Bach, S., Binder, A., Montavon, G., Klauschen, F., Müller, K.-R., & Samek, W. (2015) · PLOS ONE · doi:10.1371/journal.pone.0130140 · Cited by 4,524 · cited 25×
“We specifically use LRP as the attribution method underlying SpRAy, as this provided us with better results than other considered approaches (e.g., SmoothGrad, Integrated Gradients).”
BackgroundMethodInfluentialSelf-citationTschandl, P., Rosendahl, C., & Kittler, H. (2018) · Scientific Data · doi:10.1038/sdata.2018.161 · Cited by 3,628
Grandmaster level in StarCraft II using multi-agent reinforcement learning
Vinyals, O., Babuschkin, I., Czarnecki, W. M., Mathieu, M., Dudzik, A., Chung, J., Choi, D. H., Powell, R., Ewalds, T., Georgiev, P., et al. (2019) · Nature · doi:10.1038/s41586-019-1724-z · Cited by 3,455 · cited 1×
“Throughout the last decade, Deep Neural Networks (DNNs) have enabled impressive performance leaps in a wide range of domains, from solving classification problems [1, 2], over playing and winning games competitively [3, 4] (some in real time [5, 6]), to enabling the understanding of quantum-chemical many-body systems…”
BackgroundImproved protein structure prediction using potentials from deep learning
Senior, A. W., Evans, R., Jumper, J., Kirkpatrick, J., Sifre, L., Green, T., Qin, C., Žídek, A., Nelson, A. W. R., Bridgland, A., et al. (2020) · Nature · doi:10.1038/s41586-019-1923-7 · Cited by 3,378
Learning Important Features Through Propagating Activation Differences
Shrikumar, A., Greenside, P., & Kundaje, A. (2017) · arXiv · doi:10.48550/arxiv.1704.02685 · Cited by 2,761
Methods for interpreting and understanding deep neural networks
Montavon, G., Samek, W., & Müller, K.-R. (2018) · Digital Signal Processing · doi:10.1016/j.dsp.2017.10.011 · Cited by 2,665
Self-citationCodella, N. C. F., Gutman, D., Celebi, M. E., Helba, B., Marchetti, M. A., Dusza, S. W., Kalloo, A., Liopyris, K., Mishra, N., Kittler, H., et al. (2018) · 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018) · doi:10.1109/isbi.2018.8363547 · Cited by 2,620
Shortcut learning in deep neural networks
Geirhos, R., Jacobsen, J.-H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., & Wichmann, F. A. (2020) · Nature Machine Intelligence · doi:10.1038/s42256-020-00257-z · Cited by 2,224
European Union Regulations on Algorithmic Decision Making and a “Right to Explanation”
Goodman, B., & Flaxman, S. (2017) · AI Magazine · doi:10.1609/aimag.v38i3.2741 · Cited by 2,131 · cited 1×
“There is an increased awareness that ML models need to be interpretable to its users in order to assess the validity of the decision making of the predictor [36, 37], especially in high risk settings, such as in medical applications [38, 39, 40, 41, 42, 43].”
BackgroundCausability and explainability of artificial intelligence in medicine
Holzinger, A., Langs, G., Denk, H., Zatloukal, K., & Müller, H. (2019) · WIREs Data Mining and Knowledge Discovery · doi:10.1002/widm.1312 · Cited by 1,885 · cited 1×
“There is an increased awareness that ML models need to be interpretable to its users in order to assess the validity of the decision making of the predictor [36, 37], especially in high risk settings, such as in medical applications [38, 39, 40, 41, 42, 43].”
BackgroundExplaining Deep Neural Networks and Beyond: A Review of Methods and Applications
Samek, W., Montavon, G., Lapuschkin, S., Anders, C. J., & Muller, K.-R. (2021) · Proceedings of the IEEE · doi:10.1109/jproc.2021.3060483 · Cited by 1,477 · cited 2×
“[14, 15] for an overview) in the emerging field of eXplainable Artificial Intelligence (XAI).”
BackgroundMethodSelf-citationSmoothGrad: removing noise by adding noise
Smilkov, D., Thorat, N., Kim, B., Viégas, F., & Wattenberg, M. (2017) · arXiv · doi:10.48550/arxiv.1706.03825 · Cited by 1,467 · cited 3×
“We specifically use LRP as the attribution method underlying SpRAy, as this provided us with better results than other considered approaches (e.g., SmoothGrad, Integrated Gradients).”
MethodQuantum-chemical insights from deep tensor neural networks
Schütt, K. T., Arbabzadah, F., Chmiela, S., Müller, K. R., & Tkatchenko, A. (2017) · Nature Communications · doi:10.1038/ncomms13890 · Cited by 1,445 · cited 1×
“Throughout the last decade, Deep Neural Networks (DNNs) have enabled impressive performance leaps in a wide range of domains, from solving classification problems [1, 2], over playing and winning games competitively [3, 4] (some in real time [5, 6]), to enabling the understanding of quantum-chemical many-body systems…”
BackgroundSelf-citationExplainable AI: Interpreting, Explaining and Visualizing Deep Learning
(2019) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-28954-6 · Cited by 1,377
Explaining nonlinear classification decisions with deep taylor decomposition
Montavon, G., Lapuschkin, S., Binder, A., Samek, W., & Müller, K.-R. (2017) · Pattern Recognition · doi:10.1016/j.patcog.2016.11.008 · Cited by 1,357
Self-citationKim, B., Wattenberg, M., Gilmer, J., Cai, C., Wexler, J., Viegas, F., & Sayres, R. (2017) · arXiv · doi:10.48550/arxiv.1711.11279 · Cited by 1,240
Evaluating the Visualization of What a Deep Neural Network Has Learned
Samek, W., Binder, A., Montavon, G., Lapuschkin, S., & Muller, K.-R. (2017) · IEEE Transactions on Neural Networks and Learning Systems · doi:10.1109/tnnls.2016.2599820 · Cited by 1,164
Self-citationInterpretable Explanations of Black Boxes by Meaningful Perturbation
Fong, R. C., & Vedaldi, A. (2017) · 2017 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2017.371 · Cited by 1,097 · cited 2×
“Surrogate- and sampling-based approaches, including LIME [22], Prediction Difference Analysis [23] and Meaningful Perturbations [24] view the model as an impenetrable black box and derive local explanations via proxy models and data, at the cost of increased runtime and an approximative nature of the obtained results.”
BackgroundUnmasking Clever Hans predictors and assessing what machines really learn
Lapuschkin, S., Wäldchen, S., Binder, A., Montavon, G., Samek, W., & Müller, K.-R. (2019) · Nature Communications · doi:10.1038/s41467-019-08987-4 · Cited by 1,047
Self-citationKusner, M. J., Loftus, J. R., Russell, C., & Silva, R. (2017) · arXiv · doi:10.48550/arxiv.1703.06856 · Cited by 995 · cited 1×
“In terms of fairness, we consider the notion of counterfactual fairness [92], where a predictor is fair regarding some protected attribute A, if its decision is not affected when A is flipped to its counterfactual value.”
BackgroundOlah, C., Mordvintsev, A., & Schubert, L. (2017) · Distill · doi:10.23915/distill.00007 · Cited by 880
Right for the Right Reasons: Training Differentiable Models by Constraining their Explanations
Ross, A. S., Hughes, M. C., & Doshi-Velez, F. (2017) · Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence · doi:10.24963/ijcai.2017/371 · Cited by 515 · cited 4×
“updating the model’s parameters via some training or finetuning scheme, as is done by [66, 67, 64, 65, 50], and for AClArC.”
BackgroundMethodInfluentialVisualizing Deep Neural Network Decisions: Prediction Difference Analysis
Zintgraf, L. M., Cohen, T. S., Adel, T., & Welling, M. (2017) · arXiv · doi:10.48550/arxiv.1702.04595 · Cited by 487 · cited 2×
“Surrogate- and sampling-based approaches, including LIME [22], Prediction Difference Analysis [23] and Meaningful Perturbations [24] view the model as an impenetrable black box and derive local explanations via proxy models and data, at the cost of increased runtime and an approximative nature of the obtained results.”
BackgroundSchramowski, P., Stammer, W., Teso, S., Brugger, A., Herbert, F., Shao, X., Luigs, H.-G., Mahlein, A.-K., & Kersting, K. (2020) · Nature Machine Intelligence · doi:10.1038/s42256-020-0212-3 · Cited by 207
Lehman, J., Clune, J., Misevic, D., Adami, C., Altenberg, L., Beaulieu, J., Bentley, P. J., Bernard, S., Beslon, G., Bryson, D. M., et al. (2020) · Artificial Life · doi:10.1162/artl_a_00319 · Cited by 181 · cited 1×
“, with the help of techniques from XAI, in a surprising number of current and former state-of-the-art ML models, in part invalidating their reported (benchmark) performance as a measure of generalization capability [51, 10, 12, 50, 42, 49, 52, 53].”
MethodMorphological and molecular breast cancer profiling through explainable machine learning
Binder, A., Bockmayr, M., Hägele, M., Wienert, S., Heim, D., Hellweg, K., Ishii, M., Stenzinger, A., Hocke, A., Denkert, C., et al. (2021) · Nature Machine Intelligence · doi:10.1038/s42256-021-00303-4 · Cited by 156
Self-citationCarter, S., Armstrong, Z., Schubert, L., Johnson, I., & Olah, C. (2019) · Distill · doi:10.23915/distill.00015 · Cited by 147
How to Explain Individual Classification Decisions
(2010) · Journal of Machine Learning Research · doi:10.5555/1756006.1859912 · Cited by 54 · cited 2×
“[16, 17, 18, 19, 20, 21, 22, 23, 24]) explanations and those providing global (e.”
Method
Cited by
Explaining Deep Neural Networks and Beyond: A Review of Methods and Applications
Samek, W., Montavon, G., Lapuschkin, S., Anders, C. J., & Muller, K.-R. (2021) · Proceedings of the IEEE · doi:10.1109/jproc.2021.3060483 · Cited by 1,477
“7Or in the case of [7] deteriorate model performance, as the identified confounding feature is exclusive to the training data.”
BackgroundMethodInfluentialSelf-citationExplainable AI Methods - A Brief Overview
Holzinger, A., Saranti, A., Molnar, C., Biecek, P., & Samek, W. (2022) · Lecture Notes in Computer Science · doi:10.1007/978-3-031-04083-2_2 · Cited by 379
Self-citationFrom attribution maps to human-understandable explanations through Concept Relevance Propagation
Achtibat, R., Dreyer, M., Eisenbraun, I., Bosse, S., Wiegand, T., Samek, W., & Lapuschkin, S. (2023) · Nature Machine Intelligence · doi:10.1038/s42256-023-00711-8 · Cited by 181
“In addition, Supplementary Note 9 provides an example on how CRP can be leveraged to identify systematically learned biases in male versus female face classification and Supplementary Note 10 demonstrates the applicability of CRP to time-series data.”
BackgroundSelf-citationExplaining Machine Learning Models for Clinical Gait Analysis
Slijepcevic, D., Horst, F., Lapuschkin, S., Horsak, B., Raberger, A.-M., Kranzl, A., Samek, W., Breiteneder, C., Schöllhorn, W. I., & Zeppelzauer, M. (2022) · ACM Transactions on Computing for Healthcare · doi:10.1145/3474121 · Cited by 82
Self-citationEstablishing and evaluating trustworthy AI: overview and research challenges
Kowald, D., Scher, S., Pammer-Schindler, V., Müllner, P., Waxnegger, K., Demelius, L., Fessl, A., Toller, M., Mendoza Estrada, I. G., Šimić, I., et al. (2024) · Frontiers in Big Data · doi:10.3389/fdata.2024.1467222 · Cited by 59
Metadata
- Authors
- Anders, C. J., Weber, L., Neumann, D., Samek, W., Müller, K.-R., & Lapuschkin, S.
- Venue
- Information Fusion
- Publisher
- Elsevier BV
- Published
- 2022-01
- Volume
- 77
- Pages
- 261-295
- ISSN
- 1566-2535 (ISSN-L)
- OpenAlex
- W3114162576
- Cited by
- 119
- References
- 40
- FWCI
- 6.26
- Type
- journal article
- Language
- en
- Open access
- hybrid
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
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.