Pith. sign in

REVIEW 6 major objections 5 minor 59 references

Machine Unlearning for Robust DNNs: Attribution-Guided Partitioning and Neuron Pruning in Noisy Environments

T0 review · 6 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that attribution-based partitioning, neuron pruning, and fine-tuning on clean samples lifts noisy CIFAR-10 accuracy roughly 10 points above retraining while cutting training time by up to 47%.

desk verdict A plausible pruning-and-fine-tuning pipeline whose headline label-noise claim is untested; the core idea deserves careful review, but not acceptance as is. read the letter →

arxiv 2506.11615 v1 pith:M3RBKC6B submitted 2025-06-13 cs.LG

Deliang Jin , Gang Chen , Shuo Feng , Yufeng Ling , Haoran Zhu This is my paper · ORCID
classification cs.LG
keywords machineunlearningneuronpruningfine-tuningattributionmethodsnoise-robustlearninglabelnoiseGaussianmixturemodelsampleselection
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Drawing on the idea of machine unlearning — removing the influence of specific training samples from a trained model — this paper proposes a three-phase recipe for making an already-trained network robust to noisy data without retraining from scratch. The first phase scores each training sample by how strongly its neuron activations drive the model's output, then fits a two-component Gaussian mixture model to those score vectors to split the data into a high-quality subset and a corrupted subset, with no explicit noise-distribution assumptions. The second phase treats the inferred clean/noisy labels as targets of a linear regression on the layer's activations, producing a sensitivity score per neuron, and zeros out the top 15% of noise-sensitive neurons. The third phase fine-tunes the pruned network on the clean subset only. Across CIFAR-10 image classification and a speech-commands recognition task, the authors report that the full-model fine-tuning variant beats both the noisy initial model and standard retraining on the cleaned subset, with roughly 10 absolute points of accuracy gain on CIFAR-10 and up to 47% time savings.

What carries the argument

The load-bearing object is the per-sample attribution vector, built as the elementwise product of a layer's activations with the gradient of the network output with respect to those activations, $A^l_{m,d} = |f^l_m \odot \nabla_{f^l_m} y^d_m|$, then max-pooled across output dimensions (Eqs. 3.1–3.2). A two-component Gaussian mixture model (a soft clustering of the per-sample score vectors), fit by expectation-maximization with k-means++ initialization, converts these vectors into probabilistic clean/noisy assignments (Eqs. 3.3–3.4). A least-squares regression of those assignments on the layer activations (Eq. 3.5) produces per-neuron coefficients whose magnitudes define a sensitivity score $s_n = |T^l_n| + \lambda |u^l_n|$ (Eq. 3.6), which ranks neurons for pruning by zeroing incoming weights and biases (Eq. 3.7). A regularized fine-tuning objective over the clean subset, applied either layer-wise or to the full model (Eqs. 3.8–3.10), restores performance while keeping parameters near the pruned state.

What would settle it

Because the experiments inject noise into half of the training samples, the true clean/noisy identity of every sample is known: measure the agreement (for example, cluster purity or adjusted Rand index) between the Gaussian mixture's $\mathcal{D}_r$/ $\mathcal{D}_n$ assignment and the injected-noise mask on the 50k CIFAR-10 run. If agreement is near chance, the attribution signal is not separating clean from corrupted samples and the reported gains cannot be attributed to the claimed mechanism. A second check is to rerun the pipeline with the two cluster labels swapped, treating the presumed noisy cluster as the clean one: if that variant matches the original results, the attribution clustering is not doing the work.

Watch

Extended reading notes

Core claim

The central claim is that the influence of corrupted training samples can be removed from a trained network by locating where the noise lives, rather than by retraining the whole model. The method computes per-sample attribution vectors $A^l_{m,d} = |f^l_m \odot \nabla_{f^l_m} y^d_m|$ — the elementwise product of layer activations with the gradient of the network output with respect to those activations — max-pools them across output dimensions, and fits a two-component Gaussian mixture model so that the inferred subsets $\mathcal{D}_r$ (high-quality) and $\mathcal{D}_n$ (noise-corrupted) need no external noise assumptions. Regressing the inferred quality labels on neuron activations yields sensitivity scores $s_n = |T^l_n| + \lambda |u^l_n|$, and the top-$\alpha$ neurons are pruned by zeroing their weights and biases. Fine-tuning the pruned network on $\mathcal{D}_r$ is reported to recover and exceed the accuracy of the corrupted model: at the full 50k training scale on CIFAR-10, full-model fine-tuning reaches 80.20% accuracy versus 69.44% for the noisy initial model and 72.99% for standard retraining on the cleaned subset, and the same pattern is reported across smaller training scales, higher noise levels, and a 10-class speech-commands keyword task.

Load-bearing premise

The whole pipeline rests on one assumption: noisy and clean training samples leave separable patterns in the scores measuring how strongly each neuron responds to each sample, so the two-component Gaussian mixture really does split the data into high-quality and corrupted subsets — if the clusters do not track true sample quality, the neuron sensitivity scores and the fine-tuning data set are both built from wrong labels.

Editorial extensions

If this is right

  • Noise robustness becomes a post-training intervention: an existing model can be cleaned by partitioning, pruning, and fine-tuning, without designing noise-robust losses or estimating noise transition matrices.
  • The pruning step yields an inspectable list of noise-sensitive neurons, so the 'forgetting' of corrupted examples is localized to specific computational units rather than smeared across the network.
  • Fine-tuning for 30 epochs on the cleaned subset is reported to beat 60-epoch full retraining in both accuracy and per-epoch time, promising cheaper robustness at scale.
  • The advantage persists when training data is scarce (61.24% versus 58.28% accuracy at 12.5k samples) and generally grows with noise level, indicating the mechanism matters most where data quality is worst.
  • Results on both CIFAR-10 images and Speech Commands audio indicate the pipeline is not tied to a single modality.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • In my reading, the method does not need the Gaussian components to literally mean 'clean' and 'noisy': it needs the two clusters to separate samples whose influence helps generalization from samples whose influence hurts it, and comparing the cluster assignments against the true injected-noise mask would settle which interpretation holds.
  • A natural unstated application is defense against deliberately corrupted data: the same attribution signal that flags accidental noise could flag poisoned or adversarial examples, and the pruning step would delete their influence from the model.
  • The paper prunes only the fully connected layer of its hybrid CNN-FNN models and leaves convolutional layers untouched, so the framework's own results leave open whether the same attribution signal exists inside convolutional feature maps — extending pruning to those layers is the direct next test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

6 major / 5 minor

Summary. The paper proposes a three-phase framework, RLAP, for training noise-robust deep networks: (1) partition the training set into 'high-quality' and 'noise-corrupted' subsets by fitting a two-component Gaussian mixture model to gradient-based per-sample attribution scores; (2) prune the neurons in one FNN layer whose activations best predict the partition, using a linear regression; (3) fine-tune the pruned model on the high-quality subset. The authors report experiments on CIFAR-10 and Speech Commands, claiming large accuracy gains and reduced retraining time, including an abstract-level claim of 'approximately a 10% absolute accuracy improvement over standard retraining on CIFAR-10 with injected label noise.' The central mechanism is the assumed correspondence between one GMM component and clean data; all downstream steps depend on that partition.

Significance. If the empirical claims were fully supported, the paper would address an important practical problem: improving robustness to noisy training data without explicit noise modeling or full retraining. The proposed pipeline has intuitive appeal, and the comparison of layer-wise versus full-model fine-tuning after pruning is a reasonable design choice for an empirical study. However, the paper does not currently deliver on its stated contribution. No code or data is provided; every reported number comes from a single run without error bars or seeds; the noise levels in Table 3 are undefined; the headline label-noise experiment is absent; and the 'standard retraining' baseline is not standard retraining. These missing pieces are not cosmetic, because the central claims and the method's core assumption (GMM cluster purity) are unverified.

major comments (6)
  1. [Abstract and Section 4.1.1] The abstract claims 'approximately a 10% absolute accuracy improvement over standard retraining on CIFAR-10 with injected label noise,' but Section 4.1.1 explicitly states: 'Our test results only consider noise in features for classification tasks.' None of the reported CIFAR-10 experiments inject label noise, so the headline claim is unsupported by the evidence in the paper.
  2. [Section 4, 'Retrain Model' baseline] The retrained baseline is not standard retraining on the full noisy training set. The paper states that the retrained model 'differs from the initial model solely in the composition of the training set, where the full dataset is replaced with the refined subset Dr obtained through the data differentiation process described in Section 3.2.' Therefore, the gains in Tables 1 and 3 over the 'Retrain Model' measure the combined effect of data curation plus pruning/fine-tuning, not an improvement over standard retraining. A direct comparison against a model trained on the full noisy dataset is required.
  3. [Section 3.2.2 and Table 3] The entire method rests on the assumption that the two GMM components correspond exactly to clean versus noisy samples, with component 1 assigned to high-quality data in Eq. (3.4). The paper never validates this partition against known noise labels, and Table 3 does not define what 'level=2' through 'level=9' mean (noise type, magnitude, or proportion). If the partition is wrong, the regression labels in Eq. (3.5) and the fine-tuning set in Eq. (3.8) are meaningless. Report cluster-purity metrics (e.g., adjusted Rand index or precision/recall against the injected noise mask) and precisely specify the noise generation process for every level.
  4. [Section 3.3, Eq. (3.6) and Algorithm 1, line 19] The sensitivity score is defined as sn = |T_l^n| + lambda * |u_l^n|, but in Eq. (3.5) u_l is a scalar intercept, not a per-neuron parameter. The notation u_l^n is undefined, and the hyperparameter lambda is never specified or ablated. As written, the pruning criterion is ambiguous and not reproducible.
  5. [Tables 1-4] All results are reported for a single run without standard errors, confidence intervals, or seeds. Some differences that support the method's claims are small (e.g., Table 2, 50k F-FT: 80.20% versus 79.88% for the Standard-deviation baseline), and Table 3 shows the 'Retrain Model' sometimes being comparable to or better than F-FT (e.g., level=8: F-FT 80.20% versus Retrain 72.99% is large, but level=2: F-FT 79.34% versus Retrain 71.27% is also large; the issue is variance). Without repeated runs, none of these comparisons can be assessed statistically.
  6. [Section 3.2.2, Eq. (3.3)-(3.4)] There is a potential circularity in the learning loop: the quality labels z_m are produced by a GMM fitted to attribution scores, and the regression in Eq. (3.5) then predicts exactly those labels from activations, reinforcing the same partition. While the final test-accuracy comparison provides some external grounding, no independent evidence shows that the pruned neurons are specifically 'noise-sensitive' rather than merely correlated with the GMM's own clustering. An experiment comparing GMM-based labels against ground-truth noise masks would address this.
minor comments (5)
  1. [Section 2.2 and Eq. (3.1)] Integrated Gradients is introduced in Eq. (2.2), but the attribution used in Eq. (3.1) is simply the gradient multiplied by the activation, not an integrated-gradient computation. The relationship between these two formulations should be clarified.
  2. [Section 4.1.1] The sentence 'We compare our proposed methods with the results of the company's experiment' is unclear; please state exactly whose results are used and provide a citation or describe the baseline protocol.
  3. [Algorithm 1] The 'Require' list and step 5 both list 'Pretrained FNN model f_theta with L layers'; this duplicate line should be removed.
  4. [Throughout] There are frequent typos and formatting errors, including 'paramters', 'e fficient', 'di fferent', 'to e fficiently', and inconsistent use of math notation (e.g., 'f_theta' versus 'f_theta'). A careful proofreading pass is needed.
  5. [Section 4.1.2, Table 3] The noise-level rows are labeled 'level=2' through 'level=9' with no definition; if these correspond to increasing feature-noise magnitudes, the x-axis should be described explicitly and the noise type (e.g., Gaussian, salt-and-pepper, adversarial) should be stated.

Circularity Check

1 steps flagged · score 6.0 of 10

Neuron 'noise sensitivity' reduces to a regression fitted to the method's own GMM labels; the headline label-noise accuracy gain is not measured against standard retraining.

  1. self definitional [Section 3.2.2 Eq. (3.4) and Section 3.3 Eqs. (3.5)-(3.6)]
    "The attribution score ... is formally defined as: A^l_{m,d} = |f^l_m ⊙ ∇_{f^l_m} y^d_m|, (3.1) ... The final partitions are determined by classifying samples into high-quality data D_r = {a_m | arg max_k γ_k(a_m) = 1} and noise-corrupted data D_n = {a_m | arg max_k γ_k(a_m) = 2} ... We formulate and solve a least-squares linear regression problem to predict the sample's quality label z_m from these activations ..."

    The regression labels z_m are not ground-truth noise indicators; they are the GMM posterior assignments produced by Eq. (3.4) from attribution vectors a_m = max_d |f^l_m ⊙ ∇_{f^l_m} y^d_m| (Eqs. 3.1-3.2). The features in the same regression are the corresponding layer-l activations f^l_m. Thus s_n is the coefficient of a model fitted to a clustering of a function of the very same activations; naming the top-scoring neurons 'noise-sensitive' is a restatement of the fit, not an independent discovery. No external noise mask is used to check that component 1 is actually clean. The final test accuracy is external, so the whole pipeline is not fully circular, but the pruning-target claim itself reduces to the internal GMM labels.

full rationale

The central circular step is in the pruning criterion: quality labels z_m are generated by the GMM partition of attribution scores (Eqs. 3.1-3.4), and Eq. (3.5) regresses those same activations onto those self-generated labels; the resulting sensitivity score (Eq. 3.6) is therefore a summary of the model's own partition rather than a measured property of noise. Test-set accuracy provides an external check of the final fine-tuned model, which prevents the whole method from being vacuous, but the specific claim that pruned neurons are 'primarily influenced by noisy samples' is definitionally tied to the unvalidated cluster assignment. No author self-citation chain is involved; the cited pruning baseline [59] is external. Two further, non-circular problems are flagged per the review rule: Section 4.1.1 states 'Our test results only consider noise in features for classification tasks, but our method can easily be extended to tasks where both features and labels have noise', so the abstract's 'injected label noise' headline is not actually run; and the 'Retrain Model' baseline is defined as retraining on the GMM-selected D_r rather than on the full noisy set, so the reported accuracy gap is not an improvement over standard retraining. These issues are primarily correctness/verification failures; the circularity score is driven by the self-referential definition of noise-sensitive neurons.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method rests on four unverified modeling choices: GMM cluster equals clean data, regression coefficients equal noise sensitivity, single-layer pruning with a fixed ratio is sufficient, and the unreported hyperparameters alpha, lambda, and lambda_reg have benign values. None of these are derived or validated against ground-truth noise labels.

free parameters (4)
  • Pruning ratio alpha = 0.15
    Fixed in all experiments after preliminary ablation studies; directly controls how many neurons are removed and is central to the reported accuracy.
  • Sensitivity bias weight lambda (Eq. 3.6)
    Balances regression coefficient and bias contribution in the neuron sensitivity score; no value or tuning protocol is reported.
  • Fine-tuning regularization coefficient lambda_reg (Eqs. 3.8 to 3.10)
    Controls deviation from the pruned weights during fine-tuning; no value is reported.
  • GMM component count K = 2
    Set to 2 because the partition is binary; the choice is asserted without justification or sensitivity analysis.
assumptions (4)
  • domain assumption Noisy samples exhibit anomalous attribution patterns, and a two-component GMM on max-pooled attribution vectors separates clean from corrupted samples.
    Section 3.2.2, Eqs. 3.3 and 3.4. The paper asserts this without independent validation; component 1 is defined to be Dr by construction.
  • domain assumption Least-squares coefficients predicting GMM-derived quality labels from activations measure per-neuron sensitivity to noise, so pruning the largest scores removes noise-sensitive neurons.
    Section 3.3, Eqs. 3.5 to 3.7. No causal or identifiability argument is given, and the labels are outputs of the attribution GMM.
  • ad hoc to paper Pruning a single fixed FNN layer with alpha=0.15 is sufficient to improve robustness without harming clean performance.
    Section 4 states pruning is applied only to the FNN layer due to the contextual complexity of CNN layers; alpha is chosen by unpublished ablation.
  • standard math Expectation-Maximization for the GMM converges to a useful local optimum and k-means++ initialization is sufficient.
    Used in Section 3.2.2; standard background result for GMM fitting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Machine Unlearning for Robust DNNs: Attribution-Guided Partitioning and Neuron Pruning in Noisy Environments." pith.science (2026). https://pith.science/paper/M3RBKC6B

@misc{pith2026250611615,
  author       = {Pith},
  title        = {Pith review of: Machine Unlearning for Robust DNNs: Attribution-Guided Partitioning and Neuron Pruning in Noisy Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M3RBKC6B}},
  note         = {Machine review of arXiv:2506.11615}
}
read the original abstract

Deep neural networks (DNNs) have achieved remarkable success across diverse domains, but their performance can be severely degraded by noisy or corrupted training data. Conventional noise mitigation methods often rely on explicit assumptions about noise distributions or require extensive retraining, which can be impractical for large-scale models. Inspired by the principles of machine unlearning, we propose a novel framework that integrates attribution-guided data partitioning, discriminative neuron pruning, and targeted fine-tuning to mitigate the impact of noisy samples. Our approach employs gradient-based attribution to probabilistically distinguish high-quality examples from potentially corrupted ones without imposing restrictive assumptions on the noise. It then applies regression-based sensitivity analysis to identify and prune neurons that are most vulnerable to noise. Finally, the resulting network is fine-tuned on the high-quality data subset to efficiently recover and enhance its generalization performance. This integrated unlearning-inspired framework provides several advantages over conventional noise-robust learning approaches. Notably, it combines data-level unlearning with model-level adaptation, thereby avoiding the need for full model retraining or explicit noise modeling. We evaluate our method on representative tasks (e.g., CIFAR-10 image classification and speech recognition) under various noise levels and observe substantial gains in both accuracy and efficiency. For example, our framework achieves approximately a 10% absolute accuracy improvement over standard retraining on CIFAR-10 with injected label noise, while reducing retraining time by up to 47% in some settings. These results demonstrate the effectiveness and scalability of the proposed approach for achieving robust generalization in noisy environments.

Figures

Figures reproduced from arXiv: 2506.11615 by the authors.

Figure 3.1
Figure 3.1. The overall algorithm flow. In the data, blue symbols represent high-quality data, while [PITH_FULL_IMAGE:figures/full_fig_p008_3_1.png] view at source ↗
Figure 4.1
Figure 4.1. Representative samples misclassified by the initial model but correctly predicted by [PITH_FULL_IMAGE:figures/full_fig_p014_4_1.png] view at source ↗
Figure 4.2
Figure 4.2. Representative samples misclassified by the initial model but correctly predicted by [PITH_FULL_IMAGE:figures/full_fig_p018_4_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

59 extracted references · 37 canonical work pages

  1. [1]

    Alex Krizhevsky, Ilya Sutskever, and Geo ffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems (NeurIPS), pages 1097–1105, 2012

    work page 2012

  2. [2]

    Deep residual learning for im- age recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for im- age recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016

    2016

  3. [3]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

    arXiv 2010

  4. [4]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), pages 4171–4186, 2019

    work page 2019

  5. [5]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), pages 5998–6008, 2017

    work page 2017

  6. [6]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhari- wal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877– 1901, 2020

    1901

  7. [7]

    Speech recognition with deep recurrent neural networks

    Alex Graves, Abdel-rahman Mohamed, and Geo ffrey Hinton. Speech recognition with deep recurrent neural networks. In 2013 IEEE international conference on acoustics, speech and signal processing, pages 6645–6649. Ieee, 2013

    work page 2013

  8. [8]

    Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups

    Geo ffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdel-rahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. IEEE Signal processing magazine, 29(6):82–97, 2012

    work page 2012

Show all 59 references
  1. [9]

    Deep speech: Scaling up end-to-end speech recognition

    Awni Hannun, Carl Case, Jared Casper, Bryan Catanzaro, Greg Diamos, Erich Elsen, Ryan Prenger, Sanjeev Satheesh, Shubho Sengupta, Adam Coates, et al. Deep speech: Scaling up end-to-end speech recognition. arXiv preprint arXiv:1412.5567, 2014

    2014 arXiv

  2. [10]

    Dermatologist-level classification of skin cancer with deep neural net- works

    Andre Esteva, Brett Kuprel, Roberto A Novoa, Justin Ko, Susan M Swetter, Helen M Blau, and Sebastian Thrun. Dermatologist-level classification of skin cancer with deep neural net- works. nature, 542(7639):115–118, 2017

    2017

  3. [11]

    A survey on deep learning in medical image analysis

    Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen Awm Van Der Laak, Bram Van Ginneken, 20 and Clara I Sánchez. A survey on deep learning in medical image analysis. Medical image analysis, 42:60–88, 2017

    2017

  4. [12]

    Human-level control through deep reinforcement learning

    V olodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529–533, 2015

    2015

  5. [13]

    Mastering the game of go with deep neural networks and tree search

    David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanc- tot, et al. Mastering the game of go with deep neural networks and tree search. nature, 529(7587):484–489, 2016

    2016

  6. [14]

    Proximal policy optimization algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  7. [15]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NeurIPS), pages 2672–2680, 2014

    2014

  8. [16]

    Evasion attacks against machine learning at test time

    Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Šrndi ´c, Pavel Laskov, Giorgio Giacinto, and Fabio Roli. Evasion attacks against machine learning at test time. In Machine learning and knowledge discovery in databases: European conference, ECML pKDD 2013, p...

    2013

  9. [17]

    Classification in the presence of label noise: a survey

    Benoît Frénay and Michel Verleysen. Classification in the presence of label noise: a survey. IEEE transactions on neural networks and learning systems, 25(5):845–869, 2013

    2013

  10. [18]

    Pervasive label errors in test sets destabilize machine learning benchmarks

    Curtis G Northcutt, Anish Athalye, and Jonas Mueller. Pervasive label errors in test sets destabilize machine learning benchmarks. arXiv preprint arXiv:2103.14749, 2021

    2021 arXiv

  11. [19]

    Selfie: Refurbishing unclean samples for robust deep learning

    Hao Song, Bo Han, Yong Liu, and Masashi Sugiyama. Selfie: Refurbishing unclean samples for robust deep learning. In International Conference on Machine Learning (ICML) , pages 5907–5915, 2019

    2019

  12. [20]

    Deep learning is robust to massive label noise

    David Rolnick, Andreas Veit, Serge Belongie, and Nir Shavit. Deep learning is robust to massive label noise. arXiv preprint arXiv:1705.10694, 2017

    2017 arXiv

  13. [21]

    Mimic-iii, a freely accessible critical care database

    Alistair EW Johnson, Tom J Pollard, Lu Shen, Li-wei H Lehman, Mengling Feng, Moham- mad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and Roger G Mark. Mimic-iii, a freely accessible critical care database. Scientific data, 3(1):1–9, 2016

    2016

  14. [22]

    A closer look at memorization in deep networks

    Devansh Arpit, Stanisław Jastrz˛ ebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In International conference on machine learning, pages...

    2017

  15. [23]

    Under- standing deep learning requires rethinking generalization

    Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Under- standing deep learning requires rethinking generalization. arXiv preprint arXiv:1611.03530, 2016

    2016 arXiv

  16. [24]

    Reconciling modern machine- learning practice and the classical bias–variance trade-o ff

    Mikhail Belkin, Daniel Hsu, Siyuan Ma, and Soumik Mandal. Reconciling modern machine- learning practice and the classical bias–variance trade-o ff. Proceedings of the National Academy of Sciences, 116(32):15849–15854, 2019

    2019

  17. [25]

    Concrete problems in ai safety

    Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mané. Concrete problems in ai safety. arXiv preprint arXiv:1606.06565, 2016

    2016 arXiv

  18. [26]

    Cleannet: Transfer learning for scalable image classifier training with label noise

    Kuang-Huei Lee, Xiaodong He, Lei Zhang, and Linjun Yang. Cleannet: Transfer learning for scalable image classifier training with label noise. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5447–5456, 2018

    2018

  19. [27]

    Tods: An automated time series outlier detection system

    Kwei-Herng Lai, Daochen Zha, Guanchu Wang, Junjie Xu, Yue Zhao, Devesh Kumar, Yile Chen, Purav Zumkhawaka, Minyang Wan, Diego Martinez, et al. Tods: An automated time series outlier detection system. In Proceedings of the aaai conference on artificial intelli- gence, volume 35...

    2021

  20. [28]

    Identifying mislabeled training data

    Carla E Brodley and Mark A Friedl. Identifying mislabeled training data. Journal of artificial intelligence research, 11:131–167, 1999

    1999

  21. [29]

    Symmetric cross entropy for robust learning with noisy labels

    Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross entropy for robust learning with noisy labels. In Proceedings of the IEEE/CVF interna- tional conference on computer vision, pages 322–330, 2019

    2019

  22. [30]

    Generalized cross entropy loss for training deep neural networks with noisy labels

    Zhilu Zhang and Mert Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. Advances in neural information processing systems, 31, 2018

    2018

  23. [31]

    Dimensionality-driven learning with noisy labels

    Xingjun Ma, Yisen Wang, Michael E Houle, Shuo Zhou, Sarah Erfani, Shutao Xia, Sudan- thi Wijewickrema, and James Bailey. Dimensionality-driven learning with noisy labels. In International Conference on Machine Learning, pages 3355–3364. PMLR, 2018

    2018

  24. [32]

    Co-teaching: Robust training of deep neural networks with extremely noisy labels

    Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018

    2018

  25. [33]

    Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels

    Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In International conference on machine learning, pages 2304–2313. PMLR, 2018

    2018

  26. [34]

    Making deep neural networks robust to label noise: A loss correction approach

    Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In Proceed- ings of the IEEE conference on computer vision and pattern recognition , pages 1944–1952, 2017. 22

    1944

  27. [35]

    Learning to reweight examples for robust deep learning

    Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. InInternational conference on machine learning, pages 4334–4343. PMLR, 2018

    2018

  28. [36]

    Classification with noisy labels by importance reweighting

    Tongliang Liu and Dacheng Tao. Classification with noisy labels by importance reweighting. IEEE Transactions on pattern analysis and machine intelligence, 38(3):447–461, 2015

    2015

  29. [37]

    Are anchor points really indispensable in label-noise learning? Advances in neural information processing systems, 32, 2019

    Xiaobo Xia, Tongliang Liu, Nannan Wang, Bo Han, Chen Gong, Gang Niu, and Masashi Sugiyama. Are anchor points really indispensable in label-noise learning? Advances in neural information processing systems, 32, 2019

    2019

  30. [38]

    Learning to learn from noisy labeled data

    Junnan Li, Yongkang Wong, Qi Zhao, and Mohan S Kankanhalli. Learning to learn from noisy labeled data. In Proceedings of the IEEE /CVF conference on computer vision and pattern recognition, pages 5051–5059, 2019

    2019

  31. [39]

    Meta- weight-net: Learning an explicit mapping for sample weighting

    Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta- weight-net: Learning an explicit mapping for sample weighting. Advances in neural infor- mation processing systems, 32, 2019

    2019

  32. [40]

    Towards making systems forget with machine unlearning

    Yinzhi Cao and Junfeng Yang. Towards making systems forget with machine unlearning. In 2015 IEEE symposium on security and privacy, pages 463–480. IEEE, 2015

    2015

  33. [41]

    Machine unlearning

    Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In 2021 IEEE symposium on security and privacy (SP), pages 141–159. IEEE, 2021

    2021

  34. [42]

    Amnesiac machine learning

    Laura Graves, Vineel Nagisetty, and Vijay Ganesh. Amnesiac machine learning. In Pro- ceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 11516–11524, 2021

    2021

  35. [43]

    On the necessity of au- ditable algorithmic definitions for machine unlearning

    Anvith Thudi, Hengrui Jia, Ilia Shumailov, and Nicolas Papernot. On the necessity of au- ditable algorithmic definitions for machine unlearning. In 31st USENIX security symposium (USENIX Security 22), pages 4007–4022, 2022

    2022

  36. [44]

    Eternal sunshine of the spotless net: Selective forgetting in deep networks

    Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In Proceedings of the IEEE /CVF conference on computer vision and pattern recognition, pages 9304–9312, 2020

    2020

  37. [45]

    Axiomatic attribution for deep networks

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International conference on machine learning, pages 3319–3328. PMLR, 2017

    2017

  38. [46]

    Deep inside convolutional networks: Visualising image classification models and saliency maps

    Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013

    2013 arXiv

  39. [47]

    Learning both weights and connections for efficient neural networks

    Song Han, Je ff Pool, John Tran, and William J Dally. Learning both weights and connections for efficient neural networks. NeurIPS, 28, 2015. 23

    2015

  40. [48]

    Learning sparse neural networks through l_0 regularization

    Christos Louizos, Max Welling, and Diederik P Kingma. Learning sparse neural networks through l_0 regularization. arXiv preprint arXiv:1712.01312, 2017

    2017 arXiv

  41. [49]

    Universal language model fine-tuning for text classifi- cation

    Jeremy Howard and Sebastian Ruder. Universal language model fine-tuning for text classifi- cation. arXiv preprint arXiv:1801.06146, 2018

    2018 arXiv

  42. [50]

    Neural transfer learning for natural language processing

    Sebastian Ruder. Neural transfer learning for natural language processing. PhD thesis, NUI Galway, 2019

    2019

  43. [51]

    Zeiler and Rob Fergus

    Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In European Conference on Computer Vision (ECCV), pages 818–833. Springer, 2014

    2014

  44. [52]

    Lundberg and Su-In Lee

    Scott M. Lundberg and Su-In Lee. A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems (NeurIPS), 30:4765–4774, 2017

    2017

  45. [53]

    Axiomatic attribution for deep networks

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International Conference on Machine Learning (ICML), pages 3319–3328. PMLR, 2017

    2017

  46. [54]

    Importance estimation for neural network pruning

    Pavlo Molchanov, Arun Mallya, Stephen Tyree, Iuri Frosio, and Jan Kautz. Importance estimation for neural network pruning. In CVPR, pages 11264–11272, 2019

    2019

  47. [55]

    Faster gaze prediction with dense networks and fisher pruning

    Lucas Theis, Iryna Korshunova, Alykhan Tejani, and Ferenc Huszár. Faster gaze prediction with dense networks and fisher pruning. arXiv preprint arXiv:1801.05787, 2018

    2018 arXiv

  48. [56]

    Contributions to the mathematical theory of evolution

    Karl Pearson. Contributions to the mathematical theory of evolution. Philosophical Trans- actions of the Royal Society of London. A, 185:71–110, 1894

  49. [57]

    k-means++: The advantages of careful seeding

    David Arthur and Sergei Vassilvitskii. k-means++: The advantages of careful seeding. Tech- nical report, Stanford, 2006

    2006

  50. [58]

    Maximum likelihood from incom- plete data via the em algorithm

    Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incom- plete data via the em algorithm. Journal of the royal statistical society: series B (method- ological), 39(1):1–22, 1977

    1977

  51. [59]

    Dissecting language models: Machine unlearning via selective pruning

    Nicholas Pochinkov and Nandi Schoots. Dissecting language models: Machine unlearning via selective pruning. arXiv preprint arXiv:2403.01267, 2024. 24

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.