Pith. sign in

REVIEW 3 major objections 7 minor 81 references

Provable unlearning in topic modeling and downstream tasks

T0 review · 3 major / 7 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read The paper proves the first provable unlearning guarantees for a pre-train-then-fine-tune pipeline, using topic models; fine-tuning raises the number of documents that can be forgotten.

desk verdict A genuine first step on provable unlearning for the pretrain-fine-tune pipeline, but the central (ε,δ)-indistinguishability proof has a load-bearing anchor-word identity gap that needs fixing. read the letter →

arxiv 2411.12600 v3 pith:RMIPIQXA submitted 2024-11-19 cs.LG cs.AI

Stanley Wei , Sadhika Malladi , Sanjeev Arora , Amartya Sanyal This is my paper · ORCID
classification cs.LGcs.AI MSC 68T5068T05
keywords machineunlearningtopicmodelsdeletioncapacitydifferentialprivacyGaussianmechanismpre-trainingandfine-tuninganchorwordsprovableguarantees
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 asks whether data can be provably removed from a model that was pre-trained and then fine-tuned, without retraining the model from scratch. It works out an answer for topic models, the simplest bag-of-words language models, by designing two unlearning algorithms with formal guarantees. The first deletes documents from the pre-trained topic model using a Newton update plus calibrated Gaussian noise, at a computational cost that does not grow with the original dataset size. The second shows that after a linear classification head is tuned on top, pre-training data can be unlearned without modifying the base model, and that simpler downstream tasks allow more deletions. If correct, these are the first provable unlearning guarantees for the pre-training and fine-tuning paradigm.

What carries the argument

The central object is the row-normalized word co-occurrence matrix $\bar{Q}$, built by counting which words appear together in documents. Under the p-separability assumption, each topic has an anchor word whose row of $\bar{Q}$ is a vertex of a simplex, and every other word row is a convex combination of anchor rows with coefficients $C_i$; the learning algorithm recovers the topic-word matrix $A$ from these coefficients. For unlearning, deleting documents shifts $\bar{Q}$, and Algorithm 1 takes one projected Newton step on each $C_i$ rather than recomputing the coefficients from scratch, then rebuilds $A$ and the topic covariance $R$. Indistinguishability from retraining is obtained by adding Gaussian noise calibrated to the $\ell_2$-sensitivity of the released matrices. The downstream algorithm confines the perturbation to the head $w$, releasing only $B=Aw$, so the base model stays untouched and the sparsity of the task head improves the deletion-capacity bound.

What would settle it

A direct simulation can settle this: draw synthetic corpora from a p-separable topic model, train on $S$ and on $S \setminus S_f$, and record whether the anchor-word recovery subroutine returns the same word set whenever $|S_f|$ is at the claimed deletion capacity. If the sets differ with non-negligible probability, the shared-coordinate comparison in the indistinguishability argument does not go through as written.

Watch

Extended reading notes

Core claim

Main Result 1 (Theorem 2) is that on a corpus of $m$ documents with $r$ topics and vocabulary size $n$, Algorithm 1 performs utility-preserving $(\epsilon,\delta)$-unlearning of $m_U = \tilde{O}(m/(r^2\sqrt{nr}))$ documents from the learned topic model $(\tilde{A}, \tilde{R})$. The released model is $(\epsilon,\delta)$-indistinguishable from a model retrained without the forgotten documents, and its entrywise error from the ground-truth topic model stays bounded. Main Result 2 (Theorem 4) considers a task that depends only on a subset $T_{\mathrm{clf}}$ of topics, with $q = \min_{k \in T_{\mathrm{clf}}} \Pr_D[z=k]$. After head tuning, Algorithm 2 unlearns $\tilde{\Omega}(mq/(r\sqrt{nr}))$ pre-training documents from the fine-tuned model $B=Aw$, without touching the base model $A$; the sparser the task's topic support, the larger $q$ and the deletion capacity. The paper also shows that if the base model and the head are released separately, unlearning reverts to the harder base-model problem, so releasing only the product $Aw$ is what buys the improvement.

Load-bearing premise

The load-bearing premise is that the unlearned run and the retrained run select the same anchor words, so their coefficient matrices live in one shared coordinate system; the cited lemma only proves the two anchor sets are close, not equal.

Editorial extensions

If this is right

  • Pre-training unlearning: Algorithm 1 deletes $\tilde{O}(m/(r^2\sqrt{nr}))$ documents from a learned topic model with update cost independent of corpus size, while keeping the model within a fixed utility loss.
  • Downstream unlearning: after linear head tuning, Algorithm 2 deletes $\tilde{\Omega}(mq/(r\sqrt{nr}))$ pre-training documents, so simpler tasks with concentrated topic support permit more deletions.
  • No base modification: when the fine-tuned model is released as the single matrix $B=Aw$, unlearning happens entirely in the head, so one pre-trained base can serve many tasks with per-task deletion.
  • Release format matters: if both $A$ and $w$ are released separately, the head is just post-processing of the base and unlearning is no easier than deleting from the base model alone.
  • The guarantees are formal: each algorithm satisfies $(\epsilon,\delta)$-indistinguishability from retraining together with a bounded-utility condition, rather than relying on heuristic or empirical unlearning.

Reading between the lines

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

  • Beyond the paper, the head-only mechanism suggests a design principle for large models: task-specific adaptation narrows the released surface of the model, so deletion capacity can grow with task specialization instead of shrinking.
  • Beyond the paper, the anchor-set coincidence assumption is directly testable and could be patched by a perturbation argument that accounts for anchor re-identification; until then the indistinguishability proof rests on a coordinate-identity premise.
  • Beyond the paper, the Gaussian-mechanism dependence on $1/\sqrt{nr}$ predicts that large-vocabulary corpora support only small forget sets unless the downstream task is very simple, which could be checked empirically on real bag-of-words data.
  • Beyond the paper, transferring this idea to autoregressive language models would require replacing co-occurrence statistics with embedding geometry; a plausible first test is whether linear-probe unlearning still beats retraining when only the probe output is released.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The paper studies machine unlearning for topic models in a pretrain-fine-tune pipeline. It defines utility-preserving (epsilon, delta)-unlearning with deletion capacity, gives Algorithm 1 that performs a Newton step on the learned coefficients followed by Gaussian noise to unlearn documents from a pretrained topic model, and Algorithm 2 that unlearns pretraining documents after linear-head fine-tuning without modifying the base model. The main claims are Theorem 2: deletion capacity O~(m/(r^2 sqrt(nr))) for the base model, and Theorem 4: deletion capacity O~(m q/(r sqrt(nr))) for a downstream task depending on topics with minimum probability q; it also proves that releasing A and w separately forces base-model unlearning. The analysis builds on the Arora et al. (2012a,b) topic-model learning guarantees and a detailed perturbation analysis.

Significance. If correct, the paper provides the first provable unlearning guarantees for a pretrain-fine-tune pipeline, with explicit deletion capacity and algorithms whose unlearning cost is independent of the original dataset size. It also formalizes the intuitively appealing point that task-specific release can increase deletion capacity. The strengths include a mostly spelled-out proof strategy, appropriate use of the Gaussian mechanism and existing learning guarantees, and an honest statement of limitations. The main caveat is that the indistinguishability argument currently relies on an unproved anchor-set identity; this is a fixable but load-bearing gap.

major comments (3)
  1. [Appendix B.2, Lemmas 13 and 15] The indistinguishability proof uses C^F with two different meanings. Lemma 13 (and Lemma 11) takes C^F to be the coefficient matrix produced by the retrained run using the anchor set recovered from Q^F, whereas Lemma 15 defines C^F_i as argmin_{C in simplex} ||(Q^F_P)^T C - (Q^F_i)^T||^2, i.e., coefficients expressed in the original anchor basis P. Lemma 16 concatenates these two objects via a triangle inequality, but without P_F = P the two C^F objects are not vectors in the same coordinate system. Lemma 2 only provides O(delta/(gamma p)) closeness of the recovered anchor sets to the ground-truth vertices; it does not assert that RecoverAnchors(S) and RecoverAnchors(S \ S_f) return the same word indices. Therefore the sensitivity bound underlying the Gaussian mechanism and the (epsilon, delta)-indistinguishability statements in Theorems 2 and 4 are not fully established. Please either prove that the same anchor words are selected with high probability under the stated deletion bound, or add an explicit change-of-basis/coupling argument that handles permuted or replaced anchor words before comparing coefficients.
  2. [Definition 4] The prose says the unlearned model must be indistinguishable from the model trained on S \ S_f, but the displayed equation compares U(S_f, A(S), T(S)) with U(empty, A(S \ S_f), T(S \ S_f)). For Algorithm 1, U with an empty forget set still applies the Gaussian mechanism and projection, so the right-hand side is not the raw retrained model A(S \ S_f). The Appendix B proofs compare the unlearned output to A^F plus fresh noise; this is consistent with the equation but not with the first sentence. The definition should be repaired, for example by defining U(empty, .) to be the retraining algorithm or by stating explicitly that the guarantee is indistinguishability from the noised retrained release, because as written the formal guarantee does not exactly express the stated gold standard.
  3. [Algorithm 2 / Section 5.2] The realistic setting is described as releasing the fine-tuned model as a single matrix B = Aw in R^{n x 1}, but Algorithm 2's return statement outputs (A^S)^dagger \bar A \bar w + xi, an r-dimensional vector. If the released artifact is actually B = A^S times this vector, then the sensitivity and utility statements in Lemmas 29-31 should be formulated for A^S \tilde v; if instead the r-vector itself is the released model, then the claim that the base model is not released needs clarification, since Theorem 3 shows that releasing A and w together makes unlearning as hard as the base case. Please specify the released object and adjust the algorithm and proofs accordingly.
minor comments (7)
  1. [Appendix C.2, Lemma 32] The proof concludes 'm_U <= \tilde Omega(mq/(r^{1.5} n^{0.5}))'; this should be an upper bound on m_U with \tilde O, not a lower bound, and the deletion-capacity statement should then be written accordingly.
  2. [Theorem 7 vs. Theorem 4] The informal Theorem 4 writes the deletion capacity as c' mq/(r sqrt(nr)), while the formal Theorem 7 has an explicit dependence on epsilon and log(1/delta) inside the min; please state in the informal result that c' depends on epsilon, delta, D, and T to avoid the appearance of a mismatch.
  3. [Definition 4] The phrase 'Let m in N be a constant that depends on the topic modeling distribution D' is confusing, since m is used as the dataset size; please rephrase to distinguish the sample-size parameter from the distribution-dependent constants.
  4. [Appendix B.2, Proposition 3] Proposition 3 imposes a lower bound m_U >= Omega(sqrt(m epsilon_0/(4ar))) before replacing delta_2 with the deletion term; this lower bound is not stated in Theorems 2 or 5 and should be made explicit or shown to be without loss for the claimed capacity.
  5. [Main Result 2] The informal statement says q in [1/r, 1], but q is defined as min_{k in T_clf} Pr_D[z = k], which can be smaller than 1/r when topics are imbalanced; either assume balanced topics or state q in [1/(ar), 1].
  6. [Algorithm 2] Algorithm 2 computes \bar R via Algorithm 1 but never uses \bar R afterward; please remove it or explain its role.
  7. [Section 4.2, proof sketch of Theorem 2] The sentence 'This upper bound ensures that the anchor words are likely unchanged per Lemma 2' is not justified by Lemma 2 as stated, since Lemma 2 gives geometric closeness rather than identity of the selected anchor-word indices; this is the same issue as in the first major comment.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the unlearning bounds are derived from prior learning-theoretic guarantees, a standard Gaussian mechanism, and new perturbation analysis; no fitted quantity is disguised as a prediction.

full rationale

The paper's load-bearing unlearning claims (Theorems 2 and 4) are not circular. The learning guarantees from Arora et al. (2012a; 2012b) are imported as external theorems with stated assumptions (p-separability, sample complexity, gamma-p-robust simplex) that do not include unlearning or the deletion-capacity target; the co-authorship overlap with this paper does not make those independent prior results circular, and they are parameter-free with assumptions that do not include the target result. The unlearning analysis then derives, rather than assumes, the sensitivity bounds: Lemmas 9-16 bound the change in the row-normalized co-occurrence matrix under deletion, Lemma 4 converts this into a bound on the Newton-updated coefficient matrix, and the Gaussian mechanism (Lemma 1) supplies (epsilon, delta)-indistinguishability. Deletion capacity falls out of requiring the resulting utility error to stay below the 0.01 threshold, with the utility target being the external ground-truth matrices A-star and R-star. The downstream result similarly derives its improved capacity from the sparsity of w-star (the parameter q), not from any fitted constant renamed as a guarantee. The main proof weakness--Appendix B.2 compares coefficients expressed in the original anchor-word basis without proving that the retrained run selects the same anchor words, so Lemma 2's closeness guarantee may not justify the statement that the anchor words are likely unchanged--is a correctness/completeness gap rather than a circularity: closing it would require an additional coupling or change-of-basis argument, but no definition or fitted parameter is constructed to equal the target result by construction.

Assumptions & free parameters 0 free parameters · 9 assumptions · 0 invented entities

The central claims rest on the published learning guarantees of Arora et al. (2012a, 2012b), standard differential privacy machinery, and a set of convexity and smoothness conditions on the downstream head. No parameters are fitted to data; all bounds are derived, and the utility target is the external ground truth A*. The main unstated premise is that the original and retrained runs share the same anchor word set; that assumption is analyzed in weakest_assumption and flagged in red_flags.

assumptions (9)
  • domain assumption Assumption 1 (p-separability): the ground-truth topic matrix A* has an anchor word for every topic, with probability at least p on its own topic and zero on all others
    Section 2.1; inherited from Arora et al. 2012b; without it, learning is NP-hard and the unlearning analysis does not start.
  • domain assumption Document generation model: documents are i.i.d. draws from D over topic distributions, with L = 2 words per document without loss of generality
    Section 2.1; the i.i.d. assumption drives the concentration bound (Lemma 9) and the factorization E[Q] = A* E[XX^T] A*^T.
  • standard math Anchor recovery and coefficient error lemmas of Arora et al. 2012a, restated here as Lemmas 2, 3, 7, and 8
    Used as black-box guarantees throughout Appendix B; published peer-reviewed results from prior literature, not rederived in this paper.
  • standard math Gaussian mechanism (Lemma 1, Dwork et al. 2014)
    Standard differential privacy tool; supplies the (epsilon, delta)-indistinguishability step.
  • domain assumption Assumption 2: learning tolerance epsilon_0 <= O(1/sqrt(n r))
    Appendix B.1; technical condition on the learning precision used in the perturbation bounds.
  • domain assumption Assumption 3: every word appears with marginal probability at least epsilon_0/(4 a r), 'without loss of generality'
    Appendix B.1; imported from Arora et al. 2012b; supports the per-word occurrence lower bound F >= m epsilon_0/(4 a r) used in Lemma 10.
  • ad hoc to paper The retrained model uses the same anchor word set P as the original model
    Section 4.2 ('the anchor words are likely unchanged per Lemma 2') and Appendix B.2; Lemma 2 gives only O(delta/(gamma p))-closeness, so this implicit premise is the main proof gap.
  • domain assumption Assumptions 4 through 8: the head loss is lambda-strongly convex, minimizers have norm at most B, the loss and its gradient are Lipschitz and Hessian-Lipschitz, and inexact minimizers satisfy tau = O(1/r)
    Appendix C; standard convexity and smoothness conditions for the linear head, needed for Lemma 20 and the Newton step bound (Lemma 21).
  • domain assumption The population co-occurrence simplex is gamma p-robust (Proposition 1)
    Appendix B.1; from Arora et al. 2012b; underpins anchor stability and the Hessian minimum-eigenvalue bound (Lemma 14).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Provable unlearning in topic modeling and downstream tasks." pith.science (2026). https://pith.science/paper/RMIPIQXA

@misc{pith2026241112600,
  author       = {Pith},
  title        = {Pith review of: Provable unlearning in topic modeling and downstream tasks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RMIPIQXA}},
  note         = {Machine review of arXiv:2411.12600}
}
read the original abstract

Machine unlearning algorithms are increasingly important as legal concerns arise around the provenance of training data, but verifying the success of unlearning is often difficult. Provable guarantees for unlearning are often limited to supervised learning settings. In this paper, we provide the first theoretical guarantees for unlearning in the pre-training and fine-tuning paradigm by studying topic models, simple bag-of-words language models that can be adapted to solve downstream tasks like retrieval and classification. First, we design a provably effective unlearning algorithm for topic models that incurs a computational overhead independent of the size of the original dataset. Our analysis additionally quantifies the deletion capacity of the model -- i.e., the number of examples that can be unlearned without incurring a significant cost in model performance. Finally, we formally extend our analyses to account for adaptation to a given downstream task. In particular, we design an efficient algorithm to perform unlearning after fine-tuning the topic model via a linear head. Notably, we show that it is easier to unlearn pre-training data from models that have been fine-tuned to a particular task, and one can unlearn this data without modifying the base model.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

81 extracted references · 38 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    A practical algorithm for topic modeling with provable guarantees, 2012 a

    Sanjeev Arora, Rong Ge, Yoni Halpern, David Mimno, Ankur Moitra, David Sontag, Yichen Wu, and Michael Zhu. A practical algorithm for topic modeling with provable guarantees, 2012 a . URL https://arxiv.org/abs/1212.4777

    arXiv 2012

  3. [3]

    Learning topic models - going beyond svd, 2012 b

    Sanjeev Arora, Rong Ge, and Ankur Moitra. Learning topic models - going beyond svd, 2012 b

    work page 2012

  4. [4]

    Large image datasets: A pyrrhic win for computer vision? In 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pp.\ 1536--1546

    Abeba Birhane and Vinay Uday Prabhu. Large image datasets: A pyrrhic win for computer vision? In 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pp.\ 1536--1546. IEEE, 2021

    work page 2021

  5. [5]

    Multimodal datasets: misogyny, pornography, and malignant stereotypes

    Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. Multimodal datasets: misogyny, pornography, and malignant stereotypes. arXiv preprint arXiv:2110.01963, 2021

    arXiv 2021

  6. [6]

    Blei and John D

    David M. Blei and John D. Lafferty. Dynamic topic models. In Proceedings of the 23rd International Conference on Machine Learning, ICML '06, pp.\ 113–120, New York, NY, USA, 2006. Association for Computing Machinery. ISBN 1595933832. doi:10.1145/1143844.1143859. URL https://doi.org/10.1145/1143844.1143859

    arXiv 2006

  7. [7]

    Latent dirichlet allocation

    David M Blei, Andrew Y Ng, and Michael I Jordan. Latent dirichlet allocation. Journal of machine Learning research, 3 0 (Jan): 0 993--1022, 2003

    2003

  8. [8]

    Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S

    Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel, Jared Quincy Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stef...

    arXiv 2022

Show all 81 references
  1. [9]

    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), pp.\ 141--159. IEEE, 2021

    2021

  2. [10]

    Applications of topic models

    Jordan Boyd-Graber, Yuening Hu, David Mimno, et al. Applications of topic models. Foundations and Trends in Information Retrieval , 11 0 (2-3): 0 143--296, 2017

    2017

  3. [11]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey W...

    1901

  4. [12]

    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, pp.\ 463--480. IEEE, 2015

    2015

  5. [13]

    Extracting training data from large language models

    Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data from large language models. In 30th USENIX Security Symposium (USENIX Security 21), pp.\ ...

    2021

  6. [14]

    Extracting training data from diffusion models

    Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pp.\ 5253--5270, 2023

    2023

  7. [15]

    Langevin unlearning: A new perspective of noisy gradient descent for machine unlearning, 2024

    Eli Chien, Haoyu Wang, Ziang Chen, and Pan Li. Langevin unlearning: A new perspective of noisy gradient descent for machine unlearning, 2024. URL https://arxiv.org/abs/2401.10371

    2024

  8. [16]

    Rochelle Choenni, Ekaterina Shutova, and Robert van Rooij. Stepmothers are mean and academics are pretentious: What do pretrained language models learn about you? In Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih (eds.), Proceedings of the 2021 Confe...

    2021 doi

  9. [17]

    How fine-tuning allows for effective meta-learning

    Kurtland Chua, Qi Lei, and Jason D Lee. How fine-tuning allows for effective meta-learning. Advances in Neural Information Processing Systems, 34: 0 8871--8884, 2021

    2021

  10. [18]

    The evolution of topic modeling

    Rob Churchill and Lisa Singh. The evolution of topic modeling. ACM Comput. Surv., 2022

    2022

  11. [19]

    Maml and anil provably learn representations

    Liam Collins, Aryan Mokhtari, Sewoong Oh, and Sanjay Shakkottai. Maml and anil provably learn representations. In International Conference on Machine Learning, pp.\ 4238--4310. PMLR, 2022

    2022

  12. [20]

    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 Jill Burstein, Christy Doran, and Thamar Solorio (eds.), Proceedings of the 2019 Conference of the North A merican Chapter of...

    2019 doi

  13. [21]

    The algorithmic foundations of differential privacy

    Cynthia Dwork, Aaron Roth, et al. The algorithmic foundations of differential privacy. Foundations and Trends in Theoretical Computer Science , 9 0 (3--4): 0 211--407, 2014

    2014

  14. [22]

    Who's harry potter? approximate unlearning in llms, 2023

    Ronen Eldan and Mark Russinovich. Who's harry potter? approximate unlearning in llms, 2023. URL https://arxiv.org/abs/2310.02238

    2023 arXiv

  15. [23]

    GitHub, Inc

    DOE 1 v. GitHub, Inc. 4:22-cv-06823, N.D. Cal. 2022

    2022

  16. [24]

    OpenAI, Inc.,

    Tremblay v. OpenAI, Inc.,. 23-cv-03416-AMO, (N.D. Cal.), 2023

    2023

  17. [25]

    Regulation ( EU ) 2016/679 of the European Parliament and of the Council

    European Parliament and Council of the European Union . Regulation ( EU ) 2016/679 of the European Parliament and of the Council . URL https://data.europa.eu/eli/reg/2016/679/oj

    2016

  18. [26]

    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 Conference on Artificial Intelligence, volume 38, pp.\ 12043--12051, 2024

    2024

  19. [27]

    Erasing concepts from diffusion models

    Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 2426--2436, 2023

    2023

  20. [28]

    The pile: An 800gb dataset of diverse text for language modeling, 2020

    Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. The pile: An 800gb dataset of diverse text for language modeling, 2020. URL https://arxiv.org/abs/2101.00027

    2020 arXiv

  21. [29]

    Making ai forget you: Data deletion in machine learning

    Antonio Ginart, Melody Guan, Gregory Valiant, and James Y Zou. Making ai forget you: Data deletion in machine learning. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d Alch\' e -Buc, E. Fox, and R. Garnett (eds.), Advances in Neural Information Processing Systems, volume 32...

    2019

  22. [30]

    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, pp.\ 9304--9312, 2020

    2020

  23. [31]

    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, pp.\ 3832--3842. PMLR, 2020

    2020

  24. [32]

    Adaptive machine unlearning

    Varun Gupta, Christopher Jung, Seth Neel, Aaron Roth, Saeed Sharifi-Malvajerdi, and Chris Waites. Adaptive machine unlearning. Advances in Neural Information Processing Systems, 34: 0 16319--16330, 2021

    2021

  25. [33]

    HaoChen and Tengyu Ma

    Jeff Z. HaoChen and Tengyu Ma. A theoretical study of inductive biases in contrastive learning. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=AuEgNlEAmed

    2023

  26. [34]

    Inexact unlearning needs more careful evaluations to avoid a false sense of privacy, 2024

    Jamie Hayes, Ilia Shumailov, Eleni Triantafillou, Amr Khalifa, and Nicolas Papernot. Inexact unlearning needs more careful evaluations to avoid a false sense of privacy, 2024. URL https://arxiv.org/abs/2403.01218

    2024 arXiv

  27. [35]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll \'a r, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 16000--16009, 2022

    2022

  28. [36]

    Fantastic copyrighted beasts and how (not) to generate them

    Luxi He, Yangsibo Huang, Weijia Shi, Tinghao Xie, Haotian Liu, Yue Wang, Luke Zettlemoyer, Chiyuan Zhang, Danqi Chen, and Peter Henderson. Fantastic copyrighted beasts and how (not) to generate them. arXiv preprint arXiv:2406.14526, 2024

    2024 arXiv

  29. [37]

    Foundation models and fair use

    Peter Henderson, Xuechen Li, Dan Jurafsky, Tatsunori Hashimoto, Mark A Lemley, and Percy Liang. Foundation models and fair use. Journal of Machine Learning Research, 24 0 (400): 0 1--79, 2023

    2023

  30. [38]

    Training compute-optimal large language models

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models. arXiv preprint arXiv:2203.15556, 2022

    2022 arXiv

  31. [39]

    Probabilistic latent semantic analysis

    Thomas Hofmann et al. Probabilistic latent semantic analysis. In UAI, volume 99, pp.\ 289--296, 1999

    1999

  32. [40]

    Approximate data deletion from machine learning models, 2021

    Zachary Izzo, Mary Anne Smart, Kamalika Chaudhuri, and James Zou. Approximate data deletion from machine learning models, 2021

    2021

  33. [41]

    Knowledge unlearning for mitigating privacy risks in language models, 2023

    Joel Jang, Dongkeun Yoon, Sohee Yang, Sungmin Cha, Moontae Lee, Lajanugen Logeswaran, and Minjoon Seo. Knowledge unlearning for mitigating privacy risks in language models, 2023. URL https://openreview.net/forum?id=zAxuIJLb38

    2023

  34. [42]

    Fine-tuning can distort pretrained features and underperform out-of-distribution

    Ananya Kumar, Aditi Raghunathan, Robbie Matthew Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=UYneFzXSJWh

    2022

  35. [43]

    Towards unbounded machine unlearning

    Meghdad Kurmanji, Peter Triantafillou, Jamie Hayes, and Eleni Triantafillou. Towards unbounded machine unlearning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/forum?id=OveBaTtUAT

    2023

  36. [44]

    Predicting what you already know helps: Provable self-supervised learning

    Jason D Lee, Qi Lei, Nikunj Saunshi, and Jiacheng Zhuo. Predicting what you already know helps: Provable self-supervised learning. Advances in Neural Information Processing Systems, 34: 0 309--323, 2021

    2021

  37. [45]

    Katherine Lee, A. Cooper, Christopher Choquette-Choo, Ken Liu, Matthew Jagielski, Niloofar Mireshghallah, Lama Ahmed, James Grimmelmann, David Bau, Christopher De Sa, Fernando Delgado, Vitaly Shmatikov, Katja Filippova, Seth Neel, Miranda Bogen, Amy Cyphert, Mark Lemley, and N...

    2024

  38. [46]

    Pachinko allocation: Dag-structured mixture models of topic correlations

    Wei Li and Andrew McCallum. Pachinko allocation: Dag-structured mixture models of topic correlations. In Proceedings of the 23rd international conference on Machine learning, pp.\ 577--584, 2006

    2006

  39. [47]

    Certified minimax unlearning with generalization rates and deletion capacity

    Jiaqi Liu, Jian Lou, Zhan Qin, and Kui Ren. Certified minimax unlearning with generalization rates and deletion capacity. Advances in Neural Information Processing Systems, 36, 2024

    2024

  40. [48]

    Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing, 2021

    Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing, 2021. URL https://arxiv.org/abs/2107.13586

    2021 arXiv

  41. [49]

    A large-scale audit of dataset licensing and attribution in ai

    Shayne Longpre, Robert Mahari, Anthony Chen, Naana Obeng-Marnu, Damien Sileo, William Brannon, Niklas Muennighoff, Nathan Khazam, Jad Kabbara, Kartik Perisetla, et al. A large-scale audit of dataset licensing and attribution in ai. Nature Machine Intelligence, 6 0 (8): 0 975--...

    2024

  42. [50]

    Cost-effective retraining of machine learning models

    Ananth Mahadevan and Michael Mathioudakis. Cost-effective retraining of machine learning models. arXiv preprint arXiv:2310.04216, 2023

    2023 arXiv

  43. [51]

    Lipton, and J

    Pratyush Maini, Zhili Feng, Avi Schwarzschild, Zachary C. Lipton, and J. Zico Kolter. Tofu: A task of fictitious unlearning for llms, 2024. URL https://arxiv.org/abs/2401.06121

    2024 arXiv

  44. [52]

    Lee, Danqi Chen, and Sanjeev Arora

    Sadhika Malladi, Tianyu Gao, Eshaan Nichani, Alex Damian, Jason D. Lee, Danqi Chen, and Sanjeev Arora. Fine-tuning language models with just forward passes. In Thirty-seventh Conference on Neural Information Processing Systems, 2023 a . URL https://openreview.net/forum?id=Vota6rFhBQ

    2023

  45. [53]

    A kernel-based view of language model fine-tuning

    Sadhika Malladi, Alexander Wettig, Dingli Yu, Danqi Chen, and Sanjeev Arora. A kernel-based view of language model fine-tuning. In International Conference on Machine Learning, pp.\ 23610--23641. PMLR, 2023 b

    2023

  46. [54]

    Sources of hallucination by large language models on inference tasks

    Nick McKenna, Tianyi Li, Liang Cheng, Mohammad Javad Hosseini, Mark Johnson, and Mark Steedman. Sources of hallucination by large language models on inference tasks. arXiv preprint arXiv:2305.14552, 2023

    2023 arXiv

  47. [55]

    Rewind-to-delete: Certified machine unlearning for nonconvex functions, 2024

    Siqiao Mu and Diego Klabjan. Rewind-to-delete: Certified machine unlearning for nonconvex functions, 2024. URL https://arxiv.org/abs/2409.09778

    2024

  48. [56]

    Having beer after prayer? measuring cultural bias in large language models

    Tarek Naous, Michael Ryan, Alan Ritter, and Wei Xu. Having beer after prayer? measuring cultural bias in large language models. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Vol...

    2024 doi

  49. [57]

    Descent-to-delete: Gradient-based methods for machine unlearning

    Seth Neel, Aaron Roth, and Saeed Sharifi-Malvajerdi. Descent-to-delete: Gradient-based methods for machine unlearning. In Algorithmic Learning Theory, pp.\ 931--962. PMLR, 2021

    2021

  50. [58]

    Variational bayesian unlearning

    Quoc Phong Nguyen, Bryan Kian Hsiang Low, and Patrick Jaillet. Variational bayesian unlearning. Advances in Neural Information Processing Systems, 33: 0 16025--16036, 2020

    2020

  51. [59]

    Maxime Oquab, Timoth \'e e Darcet, Th \'e o Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Micha...

    2024

  52. [60]

    The refinedweb dataset for falcon llm: outperforming curated corpora with web data, and web data only

    Guilherme Penedo, Quentin Malartic, Daniel Hesslow, Ruxandra Cojocaru, Alessandro Cappelli, Hamza Alobeidli, Baptiste Pannier, Ebtesam Almazrouei, and Julien Launay. The refinedweb dataset for falcon llm: outperforming curated corpora with web data, and web data only. arXiv pr...

    2023 arXiv

  53. [61]

    Efficient and generalizable certified unlearning: A hessian-free recollection approach, 2024

    Xinbao Qiao, Meng Zhang, Ming Tang, and Ermin Wei. Efficient and generalizable certified unlearning: A hessian-free recollection approach, 2024. URL https://arxiv.org/abs/2404.01712

    2024 arXiv

  54. [62]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pp...

    2021

  55. [63]

    Factoring nonnegative matrices with linear programs

    Ben Recht, Christopher Re, Joel Tropp, and Victor Bittorf. Factoring nonnegative matrices with linear programs. Advances in neural information processing systems, 25, 2012

    2012

  56. [64]

    In defense of one-vs-all classification

    Ryan Rifkin and Aldebaro Klautau. In defense of one-vs-all classification. The Journal of Machine Learning Research, 5: 0 101--141, 2004

    2004

  57. [65]

    A mathematical exploration of why language models help solve downstream tasks

    Nikunj Saunshi, Sadhika Malladi, and Sanjeev Arora. A mathematical exploration of why language models help solve downstream tasks. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=vVjIW3sEc1s

    2021

  58. [66]

    Laion-5b: An open large-scale dataset for training next generation image-text models

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Informa...

    2022

  59. [67]

    Remember what you want to forget: Algorithms for machine unlearning, 2021

    Ayush Sekhari, Jayadev Acharya, Gautam Kamath, and Ananda Theertha Suresh. Remember what you want to forget: Algorithms for machine unlearning, 2021. URL https://arxiv.org/abs/2103.03279

    2021 arXiv

  60. [68]

    Smith, and Chiyuan Zhang

    Weijia Shi, Jaechan Lee, Yangsibo Huang, Sadhika Malladi, Jieyu Zhao, Ari Holtzman, Daogao Liu, Luke Zettlemoyer, Noah A. Smith, and Chiyuan Zhang. Muse: Machine unlearning six-way evaluation for language models, 2024. URL https://arxiv.org/abs/2407.06460

    2024 arXiv

  61. [69]

    Dolma: an open corpus of three trillion tokens for language model pretraining research

    Luca Soldaini, Rodney Kinney, Akshita Bhagia, Dustin Schwenk, David Atkinson, Russell Authur, Ben Bogin, Khyathi Chandu, Jennifer Dumas, Yanai Elazar, Valentin Hofmann, Ananya Jha, Sachin Kumar, Li Lucy, Xinxi Lyu, Nathan Lambert, Ian Magnusson, Jacob Morrison, Niklas Muennigh...

    2024

  62. [70]

    Diffusion art or digital forgery? investigating data replication in diffusion models

    Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 6048--6058, 2023

    2023

  63. [71]

    On the necessity of auditable algorithmic definitions for machine unlearning

    Anvith Thudi, Hengrui Jia, Ilia Shumailov, and Nicolas Papernot. On the necessity of auditable algorithmic definitions for machine unlearning. In 31st USENIX Security Symposium (USENIX Security 22), pp.\ 4007--4022, Boston, MA, August 2022. USENIX Association. ISBN 978-1-93913...

    2022

  64. [72]

    Machine unlearning via algorithmic stability

    Enayat Ullah, Tung Mai, Anup Rao, Ryan A Rossi, and Raman Arora. Machine unlearning via algorithmic stability. In Conference on Learning Theory, pp.\ 4126--4142. PMLR, 2021

    2021

  65. [73]

    Why do pretrained language models help in downstream tasks? an analysis of head and prompt tuning

    Colin Wei, Sang Michael Xie, and Tengyu Ma. Why do pretrained language models help in downstream tasks? an analysis of head and prompt tuning. Advances in Neural Information Processing Systems, 34: 0 16158--16170, 2021

    2021

  66. [74]

    Connecting pre-trained language model and downstream task via properties of representation

    Chenwei Wu, Holden Lee, and Rong Ge. Connecting pre-trained language model and downstream task via properties of representation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/forum?id=YLOJ4aKAka

    2023

  67. [75]

    First-order ANIL provably learns representations despite overparametrisation

    O g uz Kaan Y \"u ksel, Etienne Boursier, and Nicolas Flammarion. First-order ANIL provably learns representations despite overparametrisation. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=if2vRbS8Ew

    2024

  68. [76]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 11975--11986, 2023

    2023

  69. [77]

    Towards certified unlearning for deep neural networks, 2024 a

    Binchi Zhang, Yushun Dong, Tianhao Wang, and Jundong Li. Towards certified unlearning for deep neural networks, 2024 a . URL https://arxiv.org/abs/2408.00920

    2024 arXiv

  70. [78]

    Negative preference optimization: From catastrophic collapse to effective unlearning

    Ruiqi Zhang, Licong Lin, Yu Bai, and Song Mei. Negative preference optimization: From catastrophic collapse to effective unlearning. In First Conference on Language Modeling, 2024 b . URL https://openreview.net/forum?id=MXLBXjQkmb

    2024

  71. [79]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  72. [80]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  73. [81]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.