Pith. sign in

REVIEW 4 major objections 6 minor 28 references

Forget-MI: Machine Unlearning for Forgetting Multimodal Information in Healthcare Settings

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Forget-MI is a machine unlearning method that removes both the unimodal and joint representations of patient data from a trained multimodal model, reducing membership-inference attack success and forget-set performance while preserving…

desk verdict A plausible but overclaimed unlearning method: the forgetting half shows real signal, but the utility half does not survive the paper's own Table 1. read the letter →

arxiv 2506.23145 v1 pith:KWDXJVMA submitted 2025-06-29 cs.LG cs.CRcs.CV

Shahad Hardan , Darya Taratynova , Abdelmajid Essofi , Karthik Nandakumar , Mohammad Yaqub This is my paper
classification cs.LGcs.CRcs.CV
keywords machineunlearningmultimodallearninghealthcareprivacymembershipinferenceattackrighttobeforgottenchestX-rayradiologyreportsembeddings
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

The paper proposes Forget-MI, a machine unlearning method aimed at removing a patient's data from multimodal medical models that combine chest X-rays with radiology reports. The central claim is that Forget-MI can erase both the image-only, text-only, and the joint image-text representations of the forget set, while keeping the model's performance on retained patients close to the original. Using a four-part loss with small added noise, the method lowers membership-inference attack success from 1.0 to as low as 0.571 and sharply reduces AUC and F1 on the forget set, with test AUC nearly matching a model retrained from scratch. If this works, hospitals could honor deletion requests without the cost of full retraining.

What carries the argument

The mechanism is a weighted sum of four Euclidean-distance losses applied to the embedding space of a late-fusion multimodal model. Unimodal unlearning and multimodal unlearning losses pull the unlearned model's image, text, and joint embeddings away from the original model's embeddings on the forget set, while unimodal retention and multimodal retention losses keep the corresponding embeddings on a subset of the retain set close to the original. Small Gaussian noise for images and character/word-level noise for text is added to the forget set to create a forgetting neighbourhood, which the paper argues is essential because isolated data points otherwise have limited influence on the representation. The joint embedding is produced by a multimodal adaptation gate, and all distances are Euclidean.

What would settle it

Run Forget-MI with the retention losses computed on the full retained set instead of the fixed-size sequential subset, or with randomly sampled subsets over several seeds, and compare the reported test AUC/F1 and MIA with the baselines. If the test-utility advantage disappears or the MIA advantage shrinks, then the sequential-sampling choice is the load-bearing ingredient rather than the loss formulation.

Watch

Extended reading notes

Core claim

The central discovery is that multimodal unlearning is achievable in a healthcare setting by explicitly targeting both unimodal and joint embeddings. Forget-MI computes four weighted Euclidean-distance losses: two unlearning losses push the unlearned model's image, text, and joint embeddings away from the original model's embeddings on a noise-augmented forget set, and two retention losses anchor the corresponding embeddings on a sampled retain set. On an edema-classification task with chest X-rays and reports, Forget-MI achieves the lowest membership-inference attack scores among all baselines (0.571, 0.615, and 0.810 for 3%, 6%, and 10% forget sets), lowers forget-set AUC and F1, and keeps test-set AUC within about 0.01–0.07 of the retrained model. The paper also finds that small noise is necessary for forgetting neighbourhoods, and that the optimal loss weighting shifts from unlearning-focused to retention-focused as the forget fraction grows.

Load-bearing premise

The method assumes that a small subset of the retained data—whose size is fixed to match the forget set each epoch, roughly 200 samples for a 3% forget on about 6,700 records—is enough to preserve the model's knowledge of the entire retained distribution.

Editorial extensions

If this is right

  • Patient deletion requests for multimodal clinical records could be handled by updating the trained model (~5 hours) instead of retraining from scratch (~14 hours).
  • Membership-inference attackers would find forgotten patients' records nearly indistinguishable from unseen data, with attack success dropping from 1.0 to as low as 0.571.
  • Strong forgetting is reflected in the model's inability to classify the forgotten patients' data (lower AUC and F1 on the forget set), not just in attack metrics.
  • The method remains effective across forget-set sizes of 3%, 6%, and 10%, but the optimal balance of unlearning and retention losses changes with the amount forgotten.
  • Because test AUC stays close to the retrained model, the unlearned model can continue to serve the remaining patients immediately after a deletion request.

Reading between the lines

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

  • A direct test of the retention-sampling assumption: replacing the fixed-size sequential retain subset with randomly drawn subsets of the same size (or the full retained set) would show whether the reported utility preservation depends on the sampling scheme.
  • The shift in optimal loss weighting with forget-set size suggests an adaptive weighting policy for real-world deletion requests, which arrive one patient at a time rather than as a fixed batch.
  • The noise-neighbourhood idea may transfer to other privacy-sensitive applications with narrow embedding distributions, such as financial or biometric models, where a single sample has little influence on the learned representation.
  • The paper's claim of 'matching the retrained model' applies to test AUC more than to test F1; a reader should weigh whether the F1 gap (about 0.11 at 3% forget) matters for a given clinical task.
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

4 major / 6 minor

Summary. The paper proposes Forget-MI, an approximate machine-unlearning method for multimodal medical data (MIMIC-CXR chest X-rays plus radiology reports). Starting from a model trained on the full dataset, the method computes four losses: unimodal and multimodal unlearning on a noised forget set, and unimodal and multimodal retention on a sampled retain set. The authors evaluate with forget-set AUC/F1, test-set AUC/F1, and membership-inference attack (MIA), comparing against several unlearning baselines and full retraining. The central claims are that Forget-MI achieves the lowest MIA and the largest drop in forget-set utility while maintaining test-set performance comparable to both the original and the retrained models, and that it is the first multimodal unlearning method for healthcare settings.

Significance. If the claims were fully supported, this would be a timely and useful contribution: patient-level unlearning in a multimodal clinical setting is important, the four-loss formulation with input perturbations is a reasonable design, and the authors provide code and evaluate against multiple baselines. The MIA evaluation is appropriate for an unlearning paper. However, as presented, the utility-preservation claim is contradicted by the paper's own tables, the evaluation protocol prevents statistical conclusions, and the retention mechanism is not tested on the full retained data. The novelty and practical value are therefore not yet established.

major comments (4)
  1. [Table 1] The claim of "comparable performance" to the original and retrained models is not supported by Table 1. Test macro-F1 for Forget-MI is 0.250 at 3%, 0.270 at 6%, and 0.252 at 10%, versus 0.388 for the original model and 0.362/0.427/0.382 for retraining. Test AUC is also below both comparators at 6% (0.599 vs 0.677/0.702) and at 10% (0.565 vs 0.677/0.629); at 3% it is essentially tied with retrained (0.625 vs 0.626) but still below the original (0.677). Since the abstract and introduction explicitly promise test-set performance comparable to the original model and matching the retrained model, this is a load-bearing contradiction that must be addressed.
  2. [Section 4 (Evaluation Criterion)] The "best Forget-MI setting for each forgetting percentage" is selected using the same metrics that are then reported (MIA, forget-set AUC/F1, test-set AUC/F1), with no validation split, no multiple seeds, and no confidence intervals. This selection-on-the-test-set makes the reported numbers optimistically biased and prevents any statistical conclusion about the comparison with baselines or with retraining. Please add a validation-based selection procedure or report repeated-run means with error bars.
  3. [Section 3 (Implementation Details)] The retention losses (Eqs. 3-4) are computed on a temporary retain subset whose size is fixed to match the forget set at each epoch, with points "sequentially sampling points from the available retained data" (Section 3). At 3% forget, this is about 200 samples from a roughly 6,700-sample dataset. The paper therefore does not demonstrate that Forget-MI preserves knowledge of the full retained distribution; the reported test-utility drop could partly be an artifact of this small, sequentially sampled subset. The method should be evaluated with a full retain set or with retain-subset size as a controlled variable before the utility-preservation claim is made.
  4. [Section 2 (Eqs. 1-2)] The unlearning losses directly minimize the distance between the unlearned model's embeddings on the forget set and the original model's embeddings on a noisy version of the forget set. Consequently, the reported decreases in forget-set AUC and F1 are, to a substantial degree, direct effects of the optimization objective rather than independent evidence of forgetting quality. The paper should be clearer that the only partially independent forgetting metric is MIA, and should avoid presenting the forget-set utility drop as an intrinsic advantage over baselines that do not optimize the same objective.
minor comments (6)
  1. [Section 4 title] The section title "Results & Disucssions" contains a typo; it should read "Results & Discussions".
  2. [Sections 2-3 notation] The notation "Df ∈ D" should be "Df ⊂ D", and "Dr = D − Df" should be defined consistently with Df being a subset of D.
  3. [Section 3] The statement "explicitly defining Dr = D − Df is problematic" is confusing because the same definition is used at the start of Section 2; please clarify what data are actually available during unlearning and why a subset is needed.
  4. [Section 3] The text-noise perturbation is described only as "random character and word level" noise, with no quantitative hyperparameters; only the image Gaussian noise parameters are listed. Please specify the text-noise strength or state how it was chosen.
  5. [Abstract] The abstract's specific magnitudes (MIA reduced by 0.202; forget AUC/F1 reduced by 0.221 and 0.305) are not traceable to a particular comparison in Table 1; please state the reference method and the forget percentage, or report averaged values with provenance.
  6. [Section 3 and Section 4] The evaluation section lists a "distance" metric, but no distance values are reported in a table; the numbers cited in Section 4's Baseline Comparison (0.323, 0.483, etc.) appear without a clear description of how they are computed. Please provide definitions and values.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported Forget-MI metrics are selected as the best of several configurations on the same test and forget metrics they are then claimed to predict; the test-utility component of the central claim is thus partly circular.

  1. fitted input called prediction [Section 4, 'Effect of Forget Percentage' and Tables 1-2]
    "At 6%, where the MIA score is similar between Unimodal and Multimodal settings, we consider the Multimodal experiment to be outperforming due to its lower performance on the forget set."

    The loss-weighting configuration is selected by comparing the same MIA, forget-set AUC/F1, and test AUC/F1 values that are then reported as Forget-MI's performance in Table 1 and the abstract ('our approach reduces MIA by 0.202 and decreases AUC and F1 scores on the forget set by 0.221 and 0.305'). The evaluation section describes only one test set, with no validation split for choosing among the No Noise, Equal, Multimodal, Unimodal, and Retention settings. Thus the headline numbers are the 'best combination' selected on the very metrics being reported, not predictions of a fixed method; the central utility claim is a post-hoc maximum rather than a derived result.

full rationale

The paper does not rely on a load-bearing self-citation chain: references [6], [7], [17], and [25] are prior work by other groups, and no uniqueness theorem is imported from the authors' own papers. The main circular step is in evaluation: the 'best' Forget-MI configuration is chosen per forget percentage using the reported MIA, Df AUC/F1, and Dt AUC/F1 metrics themselves, so the abstract's quantitative claims are selected rather than predicted. The drop in forget-set AUC/F1 is also a direct effect of the unlearning objective (Eqs. 1-2 minimize negative distance to the original model on Df), which is expected from the method rather than an independent discovery; however, that is standard practice in unlearning evaluation and is not by itself circular. The separate issue that Table 1 shows test macro-F1 dropping from 0.388 (original) to 0.250 (Forget-MI at 3%) contradicts the 'equivalent performance' claim is a correctness problem, not a circularity, but it compounds the selection bias. Overall, the central claim of comparable test utility is partially forced by the test-set-based selection, giving a circularity score of 6.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The method introduces two substantive free parameters (loss weights and noise level) plus several smaller ones (learning rate, epochs, retain batch size). The key assumption is that negative-distance losses on noisy embeddings erase information, and the evaluation relies on an SVM-based MIA proxy. No new entities are introduced.

free parameters (6)
  • loss weights (wuu, wur, wmu, wmr) = sum to 1; best setting chosen per forget percentage (Unimodal for 3%, Multimodal for 6%, Retention for 10%); exact…
    The four loss weights in Eq. 5 are the main tuning knobs. The paper tests Equal, Multimodal, Unimodal, and Retention settings and selects the best per forget percentage using the reported evaluation metrics, so the chosen weights are effectively fitted to the test and forget sets.
  • image Gaussian noise standard deviation sigma = 0.1 (used in main comparisons)
    The authors experimented with sigma in {0.1, 0.2} and mu in {0, 0.1, 0.2}, and settled on small noise after observing that larger noise distorts performance.
  • image Gaussian noise mean mu = 0
    Chosen as part of the noise perturbation; no analysis of alternative means except a note that higher mean reduces generalization.
  • text noise level = not specified
    The paper says character and word level noise is added randomly and that the noise must be small, but no quantitative level is reported.
  • learning rate = not specified for reported runs (1e-4 or 1e-5)
    Two learning rates are listed but the final results do not state which was used.
  • number of unlearning epochs = 30
    Fixed for all runs.
assumptions (6)
  • domain assumption Euclidean distance is the appropriate distance metric for comparing embeddings in unlearning losses (Eqs. 1-4).
    The paper selects L2 distance without justification; different metrics could change the forgetting behavior.
  • ad hoc to paper Adding small Gaussian noise to images and random character/word noise to text creates a 'forgetting neighborhood' that erases patient information when the model is pushed away from the noisy embeddings.
    This is the core mechanism of the method, introduced without theoretical or empirical support beyond the final results.
  • domain assumption The multimodal adaptation gate from [23] produces a joint embedding that captures the information to be forgotten.
    The paper relies on this prior architecture without evaluation.
  • domain assumption Membership inference attack based on an SVM trained on losses of retain and test samples is a valid measure of forgetting.
    Borrowed from [17]; the paper provides no evidence that MIA success correlates with actual data erasure.
  • domain assumption The MIMIC-CXR subset with four edema classes is representative of multimodal healthcare data.
    Single dataset, single task; no evidence of generality.
  • ad hoc to paper The sum of the four losses with weights summing to 1 is a sufficient objective for unlearning.
    No ablation or theoretical justification for the additive combination.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Forget-MI: Machine Unlearning for Forgetting Multimodal Information in Healthcare Settings." pith.science (2026). https://pith.science/paper/KWDXJVMA

@misc{pith2026250623145,
  author       = {Pith},
  title        = {Pith review of: Forget-MI: Machine Unlearning for Forgetting Multimodal Information in Healthcare Settings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KWDXJVMA}},
  note         = {Machine review of arXiv:2506.23145}
}
read the original abstract

Privacy preservation in AI is crucial, especially in healthcare, where models rely on sensitive patient data. In the emerging field of machine unlearning, existing methodologies struggle to remove patient data from trained multimodal architectures, which are widely used in healthcare. We propose Forget-MI, a novel machine unlearning method for multimodal medical data, by establishing loss functions and perturbation techniques. Our approach unlearns unimodal and joint representations of the data requested to be forgotten while preserving knowledge from the remaining data and maintaining comparable performance to the original model. We evaluate our results using performance on the forget dataset, performance on the test dataset, and Membership Inference Attack (MIA), which measures the attacker's ability to distinguish the forget dataset from the training dataset. Our model outperforms the existing approaches that aim to reduce MIA and the performance on the forget dataset while keeping an equivalent performance on the test set. Specifically, our approach reduces MIA by 0.202 and decreases AUC and F1 scores on the forget set by 0.221 and 0.305, respectively. Additionally, our performance on the test set matches that of the retrained model, while allowing forgetting. Code is available at https://github.com/BioMedIA-MBZUAI/Forget-MI.git

Figures

Figures reproduced from arXiv: 2506.23145 by the authors.

Figure 1
Figure 1. The original setting contains a model trained on multimodal patient data. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Forget-MI’s Approach: The dataset has images I and text T, and is split into retain Dr and forget Df , with noise added to the Df to create D˜ f . We have two models: original Fog and unlearning Ful, and we start with Ful = Fog. Then, Ful unlearns by computing the total loss that consists of: unimodal unlearning LUU , multimodal unlearning LMU , unimodal retention LUR, and multimodal retention LMR. The unlearning lo… view at source ↗
Figure 3
Figure 3. The original model vs. best Forget-MI at each of forgetting percentages. The subplots shows the loss distribution for the forget set relative to the test set. The plots show that, unlike the original model, our model fails at recognizing the test set from the forget set, indicating successful unlearning. ment to be outperforming due to its lower performance on the forget set. One reason behind this pattern is that t… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 24 canonical work pages

  1. [1]

    Acosta, Guido J

    Julián N. Acosta, Guido J. Falcone, Pranav Rajpurkar, and Eric J. Topol. Multi- modal biomedical ai.Nature Medicine, 28(9):1773–1784, Sep 2022

    work page 2022

  2. [2]

    Scibert: Pretrained contextualized embed- dings for scientific text.CoRR, abs/1903.10676, 2019

    Iz Beltagy, Arman Cohan, and Kyle Lo. Scibert: Pretrained contextualized embed- dings for scientific text.CoRR, abs/1903.10676, 2019

    arXiv 1903

  3. [3]

    Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot

    Lucas Bourtoule, Varun Chandrasekaran, Christopher A. Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. CoRR, abs/1912.03817, 2019

    arXiv 1912

  4. [4]

    Motives for withdrawal of participation in biobanking and participants’ willingness to allow linkages of their data.European Journal of Human Genetics, 30:1–11, 11 2021

    Reinder Broekstra, Judith Aris-Meijer, Els Maeckelberghe, and Sabine Otten. Motives for withdrawal of participation in biobanking and participants’ willingness to allow linkages of their data.European Journal of Human Genetics, 30:1–11, 11 2021

    work page 2021

  5. [5]

    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, 2015

    work page 2015

  6. [6]

    Joint Modeling of Chest Radiographs and Radiology Reports for Pulmonary Edema Assessment

    Geeticka Chauhan, Ruizhi Liao, William M. Wells III, Jacob Andreas, Xin Wang, Seth J. Berkowitz, Steven Horng, Peter Szolovits, and Polina Golland. Joint modeling of chest radiographs and radiology reports for pulmonar edema assessment. CoRR, abs/2008.09884, 2020

    work page Pith review arXiv 2008

  7. [7]

    Multidelete for multimodal machine unlearning, 2024

    Jiali Cheng and Hadi Amiri. Multidelete for multimodal machine unlearning, 2024

    work page 2024

  8. [8]

    Can bad teaching induce forgetting? unlearning in deep networks using an incom- petent teacher, 2023

    Vikram S Chundawat, Ayush K Tarun, Murari Mandal, and Mohan Kankanhalli. Can bad teaching induce forgetting? unlearning in deep networks using an incom- petent teacher, 2023

    work page 2023

Show all 28 references
  1. [9]

    A guide to the california consumer privacy act of 2018.SSRN Electronic Journal, 2018

    Lydia de la Torre. A guide to the california consumer privacy act of 2018.SSRN Electronic Journal, 2018. 10 S. Hardan et al

    2018

  2. [10]

    Zero-shot machine unlearning at scale via lipschitz regularization.CoRR, abs/2402.01401, 2024

    Jack Foster, Kyle Fogarty, Stefan Schoepf, Cengiz Öztireli, and Alexandra Brin- trup. Zero-shot machine unlearning at scale via lipschitz regularization.CoRR, abs/2402.01401, 2024

    2024 arXiv

  3. [11]

    Evaluating inexact unlearning requires revisiting forgetting.CoRR, abs/2201.06640, 2022

    Shashwat Goel, Ameya Prabhu, and Ponnurangam Kumaraguru. Evaluating inexact unlearning requires revisiting forgetting.CoRR, abs/2201.06640, 2022

    2022 arXiv

  4. [12]

    Mixed-privacy forgetting in deep networks, 2021

    Aditya Golatkar, Alessandro Achille, Avinash Ravichandran, Marzia Polito, and Stefano Soatto. Mixed-privacy forgetting in deep networks, 2021

    2021

  5. [13]

    Amnesiac machine learning, 2020

    Laura Graves, Vineel Nagisetty, and Vijay Ganesh. Amnesiac machine learning, 2020

    2020

  6. [14]

    Fusion of medical imaging and electronic health records using deep learning: a systematic review and implementation guidelines.npj Digital Medicine, 3, 12 2020

    Shih-Cheng Huang, Anuj Pareek, Saeed Seyyedi, Imon Banerjee, and Matthew Lungren. Fusion of medical imaging and electronic health records using deep learning: a systematic review and implementation guidelines.npj Digital Medicine, 3, 12 2020

    2020

  7. [15]

    Mimic-CXR, a de- identified publicly available database of chest radiographs with free-text reports

    Alistair Johnson, Tom Pollard, Seth Berkowitz, Nathaniel Greenbaum, Matthew Lungren, Chih-ying Deng, Roger Mark, and Steven Horng. Mimic-CXR, a de- identified publicly available database of chest radiographs with free-text reports. Scientific Data, 6:317, 12 2019

    2019

  8. [16]

    Kravets and V

    A. Kravets and V. Namboodiri. Zero-shot class unlearning in clip with synthetic samples, 2024

    2024

  9. [17]

    Towards unbounded machine unlearning, 2023

    Meghdad Kurmanji, Peter Triantafillou, Jamie Hayes, and Eleni Triantafillou. Towards unbounded machine unlearning, 2023

    2023

  10. [18]

    Single image unlearning: Efficient machine unlearning in multimodal large language models, 2024

    Jiaqi Li, Qianshan Wei, Chuanyi Zhang, Guilin Qi, Miaozeng Du, Yongrui Chen, and Sheng Bi. Single image unlearning: Efficient machine unlearning in multimodal large language models, 2024

    2024

  11. [19]

    The EU proposal for a general data protection regulation and the roots of the ‘right to be forgotten’.Computer Law & Security Review, 29(3):229–235, 2013

    Alessandro Mantelero. The EU proposal for a general data protection regulation and the roots of the ‘right to be forgotten’.Computer Law & Security Review, 29(3):229–235, 2013

    2013

  12. [20]

    Ronak Mehta, Sourav Pal, Vikas Singh, and Sathya N. Ravi. Deep unlearning via randomized conditionally independent hessians, 2022

    2022

  13. [21]

    Deep learning models are not robust against noise in clinical text.ArXiv, abs/2108.12242, 2021

    Milad Moradi, Kathrin Blagec, and Matthias Samwald. Deep learning models are not robust against noise in clinical text.ArXiv, abs/2108.12242, 2021

    2021 arXiv

  14. [22]

    Announcement: Privacy commis- sioner seeks federal court determination on key issue for canadians’ online reputation, Oct 2018

    Office of the Privacy Commissioner of Canada. Announcement: Privacy commis- sioner seeks federal court determination on key issue for canadians’ online reputation, Oct 2018

    2018

  15. [23]

    Integrating mul- timodal information in large pretrained transformers

    Wasifur Rahman, Md Kamrul Hasan, Sangwu Lee, AmirAli Bagher Zadeh, Chengfeng Mao, Louis-Philippe Morency, and Ehsan Hoque. Integrating mul- timodal information in large pretrained transformers. In Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault, editors,Proceedi...

    2020

  16. [24]

    DeepClean: Machine unlearning on the cheap by resetting privacy sensitive weights using the fisher diagonal, 2024

    Jiaeli Shi, Najah Ghalyan, Kostis Gourgoulias, John Buford, and Sean Moran. DeepClean: Machine unlearning on the cheap by resetting privacy sensitive weights using the fisher diagonal, 2024

    2024

  17. [25]

    Are we making progress in unlearning? findings from the first NeurIPS unlearning competition, 2024

    Eleni Triantafillou, Peter Kairouz, Fabian Pedregosa, Jamie Hayes, Meghdad Kur- manji, Kairan Zhao, Vincent Dumoulin, Julio Jacques Junior, Ioannis Mitliagkas, Jun Wan, Lisheng Sun Hosoya, Sergio Escalera, Gintare Karolina Dziugaite, Peter Triantafillou, and Isabelle Guyon. Ar...

    2024

  18. [26]

    Machine unlearning: A comprehensive survey, 2024

    Weiqi Wang, Zhiyi Tian, Chenhan Zhang, and Shui Yu. Machine unlearning: A comprehensive survey, 2024. Forget-MI: Multimodal Unlearning in Healthcare 11

    2024

  19. [27]

    DeltaGrad: Rapid retraining of machine learning models

    Yinjun Wu, Edgar Dobriban, and Susan Davidson. DeltaGrad: Rapid retraining of machine learning models. In Hal Daumé III and Aarti Singh, editors,Proceedings of the 37th International Conference on Machine Learning, volume 119 ofProceedings of Machine Learning Research, pages 1...

    2020

  20. [28]

    A unified method to revoke the private data of patients in intelligent healthcare with audit to forget.Nature Communications, 14(1), 2023

    Juexiao Zhou, Haoyang Li, Xingyu Liao, Bin Zhang, Wenjia He, Zhongxiao Li, Longxi Zhou, and Xin Gao. A unified method to revoke the private data of patients in intelligent healthcare with audit to forget.Nature Communications, 14(1), 2023

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.