Pith. sign in

REVIEW 4 major objections 5 minor 50 references

NOVO: Unlearning-Compliant Vision Transformers

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A vision transformer can be made to forget any requested class on the fly, simply by removing that class's learned key token.

desk verdict NOVO is a genuinely new training-time unlearning mechanism with strong results, but the 'any request' claim outruns the tested mask distribution. read the letter →

arxiv 2507.03281 v1 pith:SVZ55E3Z submitted 2025-07-04 cs.CV

classification cs.CV
keywords machineunlearningvisiontransformerselectiveforgettingclasslearnablepromptsdeeppromptingdataprivacyzero-shot
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

NOVO is a vision transformer variant trained to be ready for future unlearning requests: every class or sub-class gets a learnable key, and forgetting is done by withholding that key at inference time, with no fine-tuning, no gradient updates, and no need for the requested data. The paper claims this is the first approach that can erase any requested class or sub-class on the fly just by naming the class indices, while keeping accuracy on the remaining classes at or above the level of a model retrained from scratch. The key bet is that a training scheme that randomly splits each mini-batch into proxy forget and retain sets, plus a drop-and-expansion trick, makes the model generalize to unlearning requests it has never seen. If right, it would turn unlearning from an expensive post-hoc procedure into a property built into the model at training time.

What carries the argument

The load-bearing object is the per-class learnable key, a small prompt token generated by a token network from a multi-hot vector marking the retain and forget class sets, concatenated to the CLS and image tokens at every transformer layer. Prediction requires both the base weights and the keys: the classifier reads the CLS token together with the retain/forget key tokens, and because the base weights are trained only in the presence of the keys, dropping a key removes the only route to that class's information. Equally important is the drop-and-expansion random mask strategy, which prevents the model from shortcutting by predicting labels from which classes happen to be in the batch, and the deep prompting scheme that re-injects the same key tokens at all layers.

What would settle it

Take the fully trained NOVO, remove all key tokens, and run a simple linear classifier on features extracted from the base weights for a forgotten class. If the classifier identifies that class's samples above chance, the claim that withholding keys irreversibly erases information from the weights is falsified.

Watch

Extended reading notes

Core claim

The central claim is that a transformer can be made unlearning-compliant by giving each class a learnable key token and training the base weights and the keys jointly, so that neither part alone can produce predictions. At inference, removing the key for a requested class irreversibly erases that class from the model's behavior: the paper reports near-zero accuracy on the forget set and near-zero membership-inference-attack scores, with retained-class accuracy slightly above a freshly retrained model in the single-class and multi-class settings tested. The architecture injects multi-hot retain/forget masks into every layer as prompt tokens, and the training objective combines cross-entropy on the retain set with a loss that drives the forget-set logits to a uniform distribution plus an inverse cross-entropy term that makes the model misclassify forget-set samples. The drop-and-expansion strategy randomly removes classes actually present in the batch and adds classes actually absent, breaking the correlation between ground-truth labels and the mask, which the paper identifies as the reason the model forgets at all when the requested classes are present.

Load-bearing premise

The whole scheme rests on the assumption that the random retain/forget masks drawn during training represent every request the model will get later, including very large or semantically correlated forget sets and repeated sequential deletions.

Editorial extensions

If this is right

  • Any class or sub-class can be erased at inference time by removing its key, so unlearning becomes a constant-time operation once the model is trained.
  • Retained-class accuracy does not degrade as the forget set grows; in the paper's CIFAR100 runs it rises from 91.8 to 96.6 when forgetting 10, 40, or 80 classes, while the comparison methods degrade substantially.
  • Membership inference attack scores land near zero on forgotten classes, indicating the forget-set samples are not distinguishable from unseen data in the unlearned model.
  • Because the keys control which classes can be predicted, NOVO doubles as a controllable classifier that only outputs the subset of classes whose keys are present.
  • Sub-class forgetting works by removing the sub-class key, though accuracy on the forget sub-class is lower than for full classes because similar sibling sub-classes remain.

Reading between the lines

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

  • The generalization guarantee is empirical: the training-time mask distribution covers subsets up to 160 classes in the paper's tests, but the full space of possible requests is exponentially larger, so an adversarial or unusual future request could fall outside the learned regime.
  • If key withdrawal applies to instances rather than classes, NOVO could be extended to instance-level unlearning by assigning each training example its own key, which the authors note as future work.
  • A risk the paper does not quantify is that an attacker with access to several keyed checkpoints might compare internal activations across keys and reconstruct removed information, so irreversibility should be tested against attackers with checkpoint access, not just the single-model membership inference attack used in the paper.
  • The same prompt-conditioning design could support continual learning by adding new classes as new keys without retraining existing keys, though the paper does not test this.
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 / 5 minor

Summary. The manuscript proposes NOVO, a vision-transformer architecture in which each class or sub-class is associated with a learnable key, and unlearning is simulated during training by splitting each mini-batch into proxy retain and forget class sets encoded as multi-hot vectors. At inference, forgetting a class is performed by removing its key, requiring no fine-tuning or access to the forget/retain sets. The paper reports experiments on CIFAR10/20/100 and TinyImageNet-200 with ViT, CAIT, and SWIN architectures across multiple resolutions, comparing against SSD, ASSD, SalUn, retrain, and other baselines. NOVO achieves near-zero forget-set accuracy in most class-level settings while retaining or exceeding baseline retain accuracy, and ablations show that the drop-and-expansion strategy, deep prompting, and the inverse cross-entropy loss are important for the reported behavior.

Significance. Conditional on the generalization claim, NOVO is an interesting new direction: it reframes unlearning as a training-time design choice rather than a post-hoc operation, and the reported consistency across architectures, datasets, and resolutions is encouraging. The paper's strengths include a clear architectural recipe, ablations that isolate the critical drop-and-expansion mechanism, and evaluation under random forget-set permutations with standard deviations in one table. The main weakness is that the central claim—that NOVO can forget any future class/sub-class request—is tested only on a limited slice of the 2^C mask space, and the 'irreversible erasure' claim rests on MIA scores and t-SNE visualizations rather than a broader security evaluation.

major comments (4)
  1. [Sec. 3.5 and Sec. 5] The central claim in the abstract and Sec. 1 that NOVO can forget 'any set of classes/sub-classes' is not established by the current experiments. The drop-and-expansion strategy in Sec. 3.5 samples ra uniformly from [0, |Ya|) and ru from [0, |Y \ Ya|), so the training distribution over masks is concentrated on mid-sized retain sets and effectively never presents extreme masks such as retaining a single class out of 100. Sec. 5 explicitly shows that without drop-and-expansion the model forgets nothing (Af = 88.6 for CIFAR100-ED), so the entire mechanism depends on this mask distribution generalizing to arbitrary future requests. The evaluations in Tables 2 and 4 cover sequential masks [0, Cf) and random masks with three seeds, but not retain sets of size below 10 (except CIFAR10 with Cf=8), not semantically structured forget sets such as all classes of one superclass, and not repeated sequential requests. Please add stress tests for these regimes or provide a formal argument that the sampled mask distribution sufficiently covers the space of possible requests.
  2. [Sec. 3.6.4, Sec. 5, Appendix C] The claim that withholding a key 'irreversibly erases information' is supported only by MIA scores in Table 1 and a t-SNE plot in Appendix B. Logistic-regression MIA is one attack family, and t-SNE inspection is qualitative; neither rules out other membership-inference variants, attribute inference, or feature-reconstruction attacks. Moreover, Appendix C reports that when the base network is frozen, NOVO achieves Af = 1.2–3.9 (Table 7), which is not complete erasure and must be reconciled with the irreversibility statement. Please specify the threat model and provide stronger evidence—for example, linear probing on hidden representations, multiple MIA variants, or a formal notion of erasure—before claiming irreversible erasure.
  3. [Table 3 and Sec. 4.2] The sub-class forgetting result does not support the strong claim of complete forgetting. In Table 3, NOVO's mean forget accuracy is 14.4%, far from the 0.0% values reported for class-level forgetting. The text acknowledges this but does not analyze why sub-class unlearning is incomplete or whether the residual 14.4% is acceptable under the paper's own definition of forgetting. Since the abstract and introduction explicitly include sub-class unlearning in the scope, this result needs a more careful treatment, and the claim should be qualified accordingly.
  4. [Tables 1, 2, 3, 5] Most comparisons are reported as single numbers without error bars or seeds. Given that the key results are retain accuracies within 1–2 points of retrain and forget accuracies of exactly 0.0, the absence of variance makes it difficult to judge whether NOVO's retain advantage is statistically significant and whether the 0.0 values are exact or rounded. The random multi-class table (Table 4, right) includes standard deviations; please provide similar statistics for the sequential, single-class, and ablation experiments.
minor comments (5)
  1. [References] References [44] and [45] appear to be the same paper ('Going deeper with image transformers'); please check the citation numbering and deduplicate.
  2. [Table 4] The caption for Table 4 merges two different experimental setups (CAIT on CIFAR100 and random multi-class forgetting with ViT) into one caption, which is confusing. Please separate the captions or clearly delineate the two blocks.
  3. [Appendix D and Fig. 6] Figure 6 in the appendix appears to duplicate Figure 3 in the main text; if they are intentionally the same, one should be removed or cross-referenced.
  4. [Sec. 3.2 and Sec. 3.5] The symbol Ya is first defined as the unique classes present in a batch and then redefined in Eq. (9) as the new retain set after drop-and-expansion; please use distinct notation to avoid confusion.
  5. [Appendix A, Table 6] The confusion matrix in Table 6 is difficult to read due to the abbreviated labels and formatting; please reformat it with clear row/column headers and class names.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the unlearning behavior is a learned conditional response evaluated on held-out requests, not an input-equivalent fit.

full rationale

The paper's chain is an architecture plus a training scheme, not a derivation that re-imports its conclusion. The forward pass (Eqs. 2-8) maps input tokens, retain mask A, and forget mask U to logits, and the training objective (Eqs. 11-15) optimizes the model to behave differently on A versus U. The central claim—on-the-fly forgetting for future requests—is tested by feeding held-out masks (sequential [0,Cf) and random selections with three seeds, Tables 2-4) that were not used to fit any parameter. The drop-and-expansion strategy (Sec. 3.5) is a data-augmentation over the mask distribution, and the ablation (Table 5) shows the model fails without it, confirming that the mechanism is learned rather than forced by construction. No parameter is fitted to the test forget sets; no load-bearing result is imported from a self-citation; no uniqueness theorem or ansatz is smuggled in via author citations. The only caveat is that the tested mask set does not cover extreme or semantically structured configurations (e.g., a single retained class, correlated forget classes), which is a generalization/robustness concern, not circularity. Therefore score 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The learnable class keys are model parameters, not newly postulated physical entities. No new particle, force, or external entity is introduced. The free parameters and domain assumptions above capture what the central claim depends on.

free parameters (3)
  • Class/sub-class keys (learnable prompts) = Not stated; learned during training
    A separate learnable key is assigned to each class or sub-class (Sec. 3.3, 3.4). These keys control whether the model can predict a class, and are the mechanism of unlearning.
  • Token networks and projection weights (phi, xi, eta1, eta2) = Not stated; about 3% of model parameters
    Additional parameters introduced to transform the multi-hot retain/forget vectors into prompt tokens (Eqs. 2-3).
  • Loss weights beta, gamma, tau = Tuned; exact values not specified
    The final objective is a weighted sum of CE, MSE, and inverse CE (Eq. 15). The paper shows ablations but does not state the chosen values in the main text.
assumptions (3)
  • ad hoc to paper The random drop-and-expansion strategy generates a sufficiently representative distribution over retain/forget masks.
    Sec. 3.5 introduces random dropping and expansion without a theoretical justification. The ablation in Sec. 5 shows the model fails to forget without this strategy, so the model's generalization rests on this empirical choice.
  • domain assumption Membership inference attack (logistic regression from [15]) is a valid proxy for whether information about the forget set remains.
    The paper uses MIA scores to support the irreversibility claim (Sec. 4.1, Tables 1, 8, 9). This assumes the MIA captures meaningful information leakage.
  • domain assumption Jointly training base weights and prompts removes discriminative class information from the base weights alone.
    Appendix C shows via t-SNE that NOVO's prompts-less features are less discriminative than vanilla ViT, but this is qualitative evidence for a strong privacy claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NOVO: Unlearning-Compliant Vision Transformers." pith.science (2026). https://pith.science/paper/SVZ55E3Z

@misc{pith2026250703281,
  author       = {Pith},
  title        = {Pith review of: NOVO: Unlearning-Compliant Vision Transformers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SVZ55E3Z}},
  note         = {Machine review of arXiv:2507.03281}
}
read the original abstract

Machine unlearning (MUL) refers to the problem of making a pre-trained model selectively forget some training instances or class(es) while retaining performance on the remaining dataset. Existing MUL research involves fine-tuning using a forget and/or retain set, making it expensive and/or impractical, and often causing performance degradation in the unlearned model. We introduce {\pname}, an unlearning-aware vision transformer-based architecture that can directly perform unlearning for future unlearning requests without any fine-tuning over the requested set. The proposed model is trained by simulating unlearning during the training process itself. It involves randomly separating class(es)/sub-class(es) present in each mini-batch into two disjoint sets: a proxy forget-set and a retain-set, and the model is optimized so that it is unable to predict the forget-set. Forgetting is achieved by withdrawing keys, making unlearning on-the-fly and avoiding performance degradation. The model is trained jointly with learnable keys and original weights, ensuring withholding a key irreversibly erases information, validated by membership inference attack scores. Extensive experiments on various datasets, architectures, and resolutions confirm {\pname}'s superiority over both fine-tuning-free and fine-tuning-based methods.

Figures

Figures reproduced from arXiv: 2507.03281 by the authors.

Figure 1
Figure 1. The overview of the proposed architecture: The model accepts multi-hot encoding [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Multi-class Forgetting on the CIFAR100 dataset (32 [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Top: Image shows the impact of prompting and to￾ken concatenation strategy for the classification. Bottom: Image shows the impact of hyperparameter. Here on the x-axis (a,b,c) is the value of hyperparameters (β, γ, τ ) 2. Prompting Strategy: In [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Left: Before unlearning of vehicles 1 (label: 18). Right: After unlearning of vehicles 1 (label: 18). B.2. After Unlearning: In [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Left: Feature extracted from vanilla ViT-B/16. Right: Feature extracted from NOVO (minus prompts). 12 [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Left: Image shows the impact of prompting and token concatenation strategy for the classification. Right: Image shows the impact of hyperparameter. Here on the x-axis (a,b,c) is the value of hyperparameters (β, γ, τ ) D. Prompting Strategy In [PITH_FULL_IMAGE:figures/…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 33 canonical work pages

  1. [1]

    https://oag.ca.gov/privacy/ccpa, 2018

    California consumer privacy act (ccpa). https://oag.ca.gov/privacy/ccpa, 2018. 1, 2

  2. [2]

    https://gdpr- info.eu/, 2018

    General data protection regulation (gdpr). https://gdpr- info.eu/, 2018. 1, 2

  3. [3]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ah- mad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 ,

  4. [4]

    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. 2

  5. [5]

    Towards making systems for- get with machine unlearning

    Yinzhi Cao and Junfeng Yang. Towards making systems for- get with machine unlearning. In 2015 IEEE Symposium on Security and Privacy, pages 463–480. IEEE, 2015. 2

  6. [6]

    Learning to unlearn: Instance-wise unlearning for pre-trained classifiers, 2024

    Sungmin Cha, Sungjun Cho, Dasol Hwang, Honglak Lee, Taesup Moon, and Moontae Lee. Learning to unlearn: Instance-wise unlearning for pre-trained classifiers, 2024. 3

  7. [7]

    Boundary unlearning: Rapid forgetting of deep net- works via shifting the decision boundary

    Min Chen, Weizhuo Gao, Gaoyang Liu, Kai Peng, and Chen Wang. Boundary unlearning: Rapid forgetting of deep net- works via shifting the decision boundary. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7766–7775, 2023. 1

  8. [8]

    Can bad teaching induce forgetting? unlearning in deep networks using an incompetent teacher

    Vikram S Chundawat, Ayush K Tarun, Murari Mandal, and Mohan Kankanhalli. Can bad teaching induce forgetting? unlearning in deep networks using an incompetent teacher. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence, 2023. 2, 5, 14, 15

Show all 50 references
  1. [9]

    Zero-shot machine unlearning

    Vikram S Chundawat, Ayush K Tarun, Murari Mandal, and Mohan Kankanhalli. Zero-shot machine unlearning. IEEE Transactions on Information Forensics and Security , 2023. 1, 2, 5, 14, 15

  2. [10]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 7

  3. [11]

    Internlm-xcomposer2: Mastering free-form text-image composition and compre- hension in vision-language large model

    Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, et al. Internlm-xcomposer2: Mastering free-form text-image composition and compre- hension in vision-language large model. arXiv preprint arXiv:2401.1...

  4. [12]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint a...

  5. [13]

    Challenging forgets: Unveiling the worst-case forget sets in machine unlearning

    Chongyu Fan, Jiancheng Liu, Alfred Hero, and Sijia Liu. Challenging forgets: Unveiling the worst-case forget sets in machine unlearning. In European Conference on Computer Vision, pages 278–297. Springer, 2024. 1

  6. [14]

    Salun: Empowering machine un- learning via gradient-based weight saliency in both image classification and generation

    Chongyu Fan, Jiancheng Liu, Yihua Zhang, Eric Wong, Den- nis Wei, and Sijia Liu. Salun: Empowering machine un- learning via gradient-based weight saliency in both image classification and generation. In The Twelfth International Conference on Learning Representations, 2024. 3, 5, 7

  7. [15]

    Fast machine unlearning without retraining through selective synaptic dampening

    Jack Foster, Stefan Schoepf, and Alexandra Brintrup. Fast machine unlearning without retraining through selective synaptic dampening. In Proceedings of the AAAI Confer- ence on Artificial Intelligence, pages 12043–12051, 2024. 2, 3, 5, 6, 7, 8, 14, 15

  8. [16]

    How to train vision transformer on small-scale datasets? In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022

    Hanan Gani, Muzammal Naseer, and Mohammad Yaqub. How to train vision transformer on small-scale datasets? In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022. BMV A Press, 2022. 7

  9. [17]

    Mixed-privacy for- getting in deep networks

    Aditya Golatkar, Alessandro Achille, Avinash Ravichan- dran, Marzia Polito, and Stefano Soatto. Mixed-privacy for- getting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 792–801, 2021. 1

  10. [18]

    Amne- siac machine learning

    Laura Graves, Vineel Nagisetty, and Vijay Ganesh. Amne- siac machine learning. In Proceedings of the AAAI Confer- ence on Artificial Intelligence, pages 11516–11524, 2021. 2, 5, 14, 15

  11. [19]

    Certified data removal from machine learning models

    Chuan Guo, Tom Goldstein, Awni Hannun, and Laurens Van Der Maaten. Certified data removal from machine learning models. In International Conference on Machine Learning, pages 3832–3842. PMLR, 2020

  12. [20]

    Towards independence criterion in ma- chine unlearning of features and labels

    Ling Han, Nanqing Luo, Hao Huang, Jing Chen, and Mary- Anne Hartley. Towards independence criterion in ma- chine unlearning of features and labels. arXiv preprint arXiv:2403.08124, 2024. 1

  13. [21]

    Model sparsity can simplify machine unlearning, 2024

    Jinghan Jia, Jiancheng Liu, Parikshit Ram, Yuguang Yao, Gaowen Liu, Yang Liu, Pranay Sharma, and Sijia Liu. Model sparsity can simplify machine unlearning, 2024. 2, 3

  14. [22]

    Vi- sual prompt tuning

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Vi- sual prompt tuning. In European Conference on Computer Vision (ECCV), 2022. 3

  15. [23]

    Layer attack unlearning: Fast and accurate machine unlearning via layer level attack and knowledge distillation

    Hyunjune Kim, Sangyong Lee, and Simon S Woo. Layer attack unlearning: Fast and accurate machine unlearning via layer level attack and knowledge distillation. In Proceed- ings of the AAAI Conference on Artificial Intelligence, pages 21241–21248, 2024. 1

  16. [24]

    Overcoming catastrophic forgetting in neu- ral networks

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska- Barwinska, et al. Overcoming catastrophic forgetting in neu- ral networks. Proceedings of the national academy of sc...

  17. [25]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5, 6, 7, 11, 12

  18. [26]

    Towards unbounded machine unlearn- ing

    Meghdad Kurmanji, Peter Triantafillou, Jamie Hayes, and Eleni Triantafillou. Towards unbounded machine unlearn- ing. Advances in neural information processing systems, 36,

  19. [27]

    Ya Le and Xuan S. Yang. Tiny imagenet visual recognition challenge. 2015. 5

  20. [28]

    Blip-2: Bootstrapping language-image pre-training with 9 frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with 9 frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 1

  21. [29]

    Erm-ktp: Knowledge-level machine un- learning via knowledge transfer

    Shen Lin, Xiaoyu Zhang, Chenyang Chen, Xiaofeng Chen, and Willy Susilo. Erm-ktp: Knowledge-level machine un- learning via knowledge transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20147–20155, 2023. 1, 3, 7

  22. [30]

    Continual learning and private unlearning

    Bo Liu, Qiang Liu, and Peter Stone. Continual learning and private unlearning. In Conference on Lifelong Learning Agents, pages 243–254. PMLR, 2022. 1

  23. [31]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1

  24. [32]

    Model spar- sity can simplify machine unlearning

    Jiancheng Liu, Parikshit Ram, Yuguang Yao, Gaowen Liu, Yang Liu, PRANAY SHARMA, Sijia Liu, et al. Model spar- sity can simplify machine unlearning. Advances in Neural Information Processing Systems, 36, 2024. 1

  25. [33]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021. 1, 2, 3, 5, 7

  26. [34]

    Fair machine unlearning: Data removal while mitigating disparities

    Alex Oesterling, Jiaqi Ma, Flavio Calmon, and Himabindu Lakkaraju. Fair machine unlearning: Data removal while mitigating disparities. In International Conference on Arti- ficial Intelligence and Statistics , pages 3736–3744. PMLR,

  27. [35]

    Partially blinded unlearning: Class unlearning for deep networks a bayesian perspective

    Subhodip Panda, Shashwat Sourav, et al. Partially blinded unlearning: Class unlearning for deep networks a bayesian perspective. arXiv preprint arXiv:2403.16246, 2024. 1

  28. [36]

    amnesia–towards machine learning mod- els that can forget user data very fast

    Sebastian Schelter. amnesia–towards machine learning mod- els that can forget user data very fast. In 1st International Workshop on Applied AI for Database Systems and Applica- tions (AIDB19), 2019. 1

  29. [37]

    Parameter-tuning-free data entry error unlearning with adap- tive selective synaptic dampening, 2024

    Stefan Schoepf, Jack Foster, and Alexandra Brintrup. Parameter-tuning-free data entry error unlearning with adap- tive selective synaptic dampening, 2024. 2, 3, 5, 7, 14, 15

  30. [38]

    Feder- ated unlearning: a perspective of stability and fairness.arXiv preprint arXiv:2402.01276, 2024

    Jiaqi Shao, Tao Lin, Xuanyu Cao, and Bing Luo. Feder- ated unlearning: a perspective of stability and fairness.arXiv preprint arXiv:2402.01276, 2024. 1

  31. [39]

    Label- agnostic forgetting: A supervision-free unlearning in deep models

    Shaofei Shen, Chenhao Zhang, Yawen Zhao, Alina Bialkowski, Weitong Tony Chen, and Miao Xu. Label- agnostic forgetting: A supervision-free unlearning in deep models. arXiv preprint arXiv:2404.00506, 2024

  32. [40]

    Deep regression unlearning

    Ayush K Tarun, Vikram S Chundawat, Murari Mandal, and Mohan Kankanhalli. Deep regression unlearning. arXiv preprint arXiv:2210.08196, 2022

  33. [41]

    Deep regression unlearn- ing

    Ayush Kumar Tarun, Vikram Singh Chundawat, Murari Mandal, and Mohan Kankanhalli. Deep regression unlearn- ing. In International Conference on Machine Learning , pages 33921–33939. PMLR, 2023

  34. [42]

    Fast yet effective machine unlearning

    Ayush K Tarun, Vikram S Chundawat, Murari Mandal, and Mohan Kankanhalli. Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Sys- tems, 2023. 1, 2, 3, 5

  35. [43]

    Unrolling sgd: Understanding factors in- fluencing machine unlearning

    Anvith Thudi, Gabriel Deza, Varun Chandrasekaran, and Nicolas Papernot. Unrolling sgd: Understanding factors in- fluencing machine unlearning. In 2022 IEEE 7th European Symposium on Security and Privacy (EuroS&P), pages 303–

  36. [44]

    Going deeper with im- age transformers

    Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Herv´e J´egou. Going deeper with im- age transformers. CoRR, abs/2103.17239, 2021. 3, 5, 7

  37. [45]

    Going deeper with im- age transformers

    Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Herv´e J´egou. Going deeper with im- age transformers. In Proceedings of the IEEE/CVF interna- tional conference on computer vision, pages 32–42, 2021. 2, 7

  38. [46]

    Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023. 1

  39. [47]

    Visualizing data using t-sne

    Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579–2605, 2008. 11

  40. [48]

    A comparative study of open-source large language models, gpt-4 and claude 2: Multiple-choice test taking in nephrology

    Sean Wu, Michael Koo, Lesley Blum, Andy Black, Liyo Kao, Fabien Scalzo, and Ira Kurtz. A comparative study of open-source large language models, gpt-4 and claude 2: Multiple-choice test taking in nephrology. arXiv preprint arXiv:2308.04709, 2023. 1

  41. [49]

    Delta- grad: Rapid retraining of machine learning models

    Yinjun Wu, Edgar Dobriban, and Susan Davidson. Delta- grad: Rapid retraining of machine learning models. In In- ternational Conference on Machine Learning, pages 10355– 10366. PMLR, 2020. 1

  42. [50]

    Arcane: An efficient architecture for exact machine unlearning

    Haonan Yan, Xiaoguang Li, Ziyao Guo, Hui Li, Fenghua Li, and Xiaodong Lin. Arcane: An efficient architecture for exact machine unlearning. In IJCAI, page 19, 2022. 1 10 A. What are forget classes misclassified as? Table 6. Confusion matrix 0 (aq mml) 8 (crnvr) 14 (ppl) 19 (vh ...

Pith tools

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