acceptodds
Under review as a conference paper at ICLR 2027

Retain-Forget Limits and Tradeoffs in Machine Unlearning with Partial Information

Abstract

Machine unlearning is commonly framed as an approximation to retraining on the retained data. While prior literature typically characterizes unlearning as a tradeoff between computational efficiency and accuracy, this work reframes the task as an optimization problem constrained by partial information about the training set, such as the original model and local Hessian of the training loss, but not the full retained data. We formalize the resulting ambiguity through the set of feasible datasets consistent with the same partial information, and define retain-set excess risk and forget-set unlearning risk, which measure the gap between an unlearned model and the retraining oracle on the retained and forgetting data respectively. For least-squares regression, we decompose both risks into algorithm-dependent fitting terms and algorithm-independent ambiguity floors. We then construct a two-layer ReLU example, showing that both ambiguity floors scale as , where is the forget-set size, even when the algorithm utilizes the original model, the forget set, and local Hessian. We further show that the retain and forget fitting objectives need not share a minimizer, yielding a Pareto tradeoff between the two goals. Finally, for representative first- and second-order unlearning methods, we derive matching risk upper bounds under suitable regularity conditions and analyze their tradeoff between retain and forget.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Machine Unlearning, Multi-Objective Optimization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_yQW5OpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies machine unlearning when the algorithm has only partial information about the training data, such as the original model, the forget set, and the local Hessian, rather than the full retained set. It defines the set of datasets consistent with this information, along with retain and forget risks relative to the retrained model. It shows that for least-squares regression these risks have irreducible floors, and gives a two-layer ReLU example where both scale as Ω(k).

Critical strengths
  1. Introduces an interesting concept: partial information in machine unlearning. Without assuming access to the full dataset, the problem becomes much harder, and the paper gives good insight into this.
  2. The definition captures the main issue. It uses a “softer” notion of unlearning that only preserves functionality: keep performance on the retain set and agree with the retrained network on the forget set, rather than matching the distribution of the retrained solution.
  3. Shows that with partial information, even these weaker unlearning goals are not necessarily achievable, and supports this with a hard example.
Critical weaknesses
  1. No experiments on real unlearning tasks to support the claims. It is unclear whether the assumptions and conclusions hold when forgetting real data.
  2. The analysis is limited to regression.

Neither of these is crucial in my view.

Questions for authors
  1. Do the authors believe these results extend to classification losses such as cross-entropy, or does the current foundation depend heavily on the regression task?
  2. Do the authors believe these results hold in real-data unlearning scenarios? Or, when encountering real-world datasets and models, will it affect the results for better or worse?
1 comment
AuthorsOpenReview ↗

Response to Reviewer yQW5

We thank Reviewer yQW5 for the positive assessment and helpful comments. We respond to the comments below.

Q1 and W2: Extension to binary classification with cross-entropy loss

We thank the reviewer for this insightful question. To address this question, we have added Appendix G, which extends the retain and forget risks decompositions to binary classification with cross-entropy loss. Specifically, Propositions 6 and 7 express the risks through algorithm-dependent KL fitting terms and algorithm-independent entropy terms, analogous to Propositions 1 and 2. The fitting terms measure agreement with the conditional mean labels and the conditional mean retrained predictions, respectively. Their non-negativity also yields lower-bound inequalities analogous to Corollary 1.

Q2 and W1: Extension to real-data unlearning scenarios

We agree with the reviewer that empirically assessing the practical implications of our results on real-world datasets is an important direction for future work. Our intuition is that similar information limitations may also hinder unlearning in real-data settings. As illustrated in Figure 1, the difficulty arises from the unlearner’s limited information about the retained-data loss and its curvature outside a local neighborhood of the original model. In practice, the high dimensionality of the parameter space and the complexity of real-world data distributions may further complicate inferring the loss landscape from these local observations.

Reviewer_8mZ7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This work formulates unlearning as a problem of producing a model as a similar as possible to the retrained model under partial information. The author assume access to only partial information about the data and they show that a retrained and forget risks are always bounded from below (irreducible), creating a fundamental impossibility of obtaining a model that is perfectly the same as the retrained model.

The authors establish that under partial information, a set of multiple feasible datasets contains fundamental ambiguity such that the learner can never pinpoint the exact dataset the retrained model would be generated by training on. This means that the unlearner is chasing multiple retrianed models at once, hence having irreducible variance in the best achievable result.

The authors give the example of varying labels for the same retrained data and decompose the retain and forget risks under this partial information regime into terms that are unlearning-dependent (reducible) which are objective of unlearning, and unlearning-independent which stem from the fundamental ambiguity created by the lack of information.

Critical strengths

S1. The direction this work is aiming at is both interesting and promising. We don't have a lot of works characterising the limits of unlearning and establishing what is achievable and what is not.

S2. The paper is easy to follow. Most notation is clear and the exposition flows nicely. The paper is well written and presents clear motivation, contribution, and treatment and explanation of the results.

Critical weaknesses

I have critical concerns regarding what the paper actually achieves. It seems to me that the paper creates a superficial problem and analyses it. The paper neither gives a solution to its created problem nor analyses an actual problem of practical importance, making me recommend rejection in its current form. I clarify this position in the comments below.

W1 [CRITICAL]. The paper assumes access to only partial information without clear specification of the practicality of that assumption in practice:

  • i) The authors cite papers proposing unlearning methods under partial information (e.g., Guo et al. 2020, Golatkar et al. 2020, Izzo et al. 2021). However, those cited works do not restrict unlearning to a partial-information style, they simply use partial information to induce unlearning from a practical point of view. In other words, they argue that the partial information they use suffices to induce unlearning. Citing them to motivate a new formulation of unlearning under partial information as if it is the norm in unlearning practices inflates the problem. What would be an interesting highlight is to investigate the limits those methods can best achieve in approximating the retrained model in practical experimental settings.
  • ii) The paper does not tell us what is the required partial information to remove the ambiguities of interest in unlearning. One doesn't have to have complete information about the data generating distribution for example to be able to perform perfect unlearning. Having partial information is fine, but what is the sufficient partial information to alleviate their raised bounds? Discussing and explaining this is fundamental to this work, and needed. I understand that this is stated as a future direction in the conclusion, but I see the work without that discussion quite insignificant.
  • iii) In practice, the unlearner has access to the full training dataset of retained and to-forget data of both features and labels. Is this considered partial information? This is the most practical case and it is not discussed in this work.
  • iv) The definition of partial information given in this work is a set of real-valued scalar functions. It is not clear how that encompasses the case when the available partial information is the data without labels or the loss function local curvature. I think the definition of the class of functions F=f:D→R\mathcal{F} = \\{f : \mathcal{D} \to \mathbb{R}\\} requires a revisit to fit the types of partial information considered in the paper.
  • v) Section 3 only deals with ambiguity related to lack of labels. As I pointed above, in practice the labels are almost always available. The reader would benefit a lot from practical unlearning scenarios in which this limitation holds. Moreover, what about other partial information cases? For example, what about the lack of one or multiple features of the data while having access to the labels? What about using only a subset of the retained or to-forget data instead of the whole datasets? Or, using a proxy of retained or to-forget data distributed according to the same distribution as those but not used for training (e.g., one is trying to forget all peacock images in CIFAR100 but has access to only peacock images from the test split not the training split)? If the focus is only on special partial information such as lack of labels or accessing only local second-order approximations of the retained loss (as is the case in sections 4 and 5), then that has to appear in the title as well as the abstract of this work. The work currently does not reflect a general treatment of all partial information types.
  • vi) The retained and forget risks are unjustifiably of cumulative loss. Why are not they averaged over the number of data points which is the practical formulation? In that case, the limits in Theorem 1 would not depend on the size of the forget set, deeming some of the central arguments of the work lapse.

W2. This work focuses on approximating the retrained model in terms of its outputs. This is stated in lines 302-303.

Intuitively, the overall target of unlearning, approximating the retrained model, can be interpreted in two ways: (i) matching the empirical loss of the retrained model on the retained dataset; (ii) matching the prediction of the retrained model on the forgetting data to prevent privacy leakage.

However, this is not the most common definition of approximate unlearning in the literature. Researchers are interested in approximating the retrained model in the parameter space rather than the output space. This seems to be implicitly targeted in the Toy Non-Linear Example starting in line 340 (particularly the equations in line 356), but it is deeply vague and confusing what the authors mean by unlearning (referred to at times as "exact", e.g., in line 620). For this matter, it is important to note that it is uncommon in the unlearning literature to target one specific retrained model, as that is an infeasible target in practice. The common target is statistical indistinguishability in the parameter space which accommodates stochasticity in what the unlearning algorithm actually achieves (cf. Guo et al. 2020). In other words, not having exactly one retrained model for a given unlearning problem is technically fine as far as the indistinguishability guarantees are concerned. This requires careful treatment from the authors in order to pinpoint the actual scale of the problem induced by the ambiguity in the dataset given the partial information.


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.

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.

Zachary Izzo, Mary Anne Smart, Kamalika Chaudhuri, and James Zou. Approximate data deletion from machine learning models. In International conference on artificial intelligence and statistics, pp. 2008–2016. PMLR, 2021.

Questions for authors

Q1. Please highlight the practicality of partial information as defined in the introduction (i.e., a set of real-valued scalar functions) in real-life applications of machine unlearning. (W1 iv)

Q2. Please include a discussion, even if intuitive, of what can be sufficient partial information for approximating a model retrained from scratch in parameter (or output thereof). This is to appreciate the treatment of other types of partial information introduced in this work and see how much that deviates from the perfect scenario. (W1 ii)

Q3. Please revisit the definition of partial information and the functions extracting information from data in this work. It does not seem that, for instance, a Hessian matrix satisfies the current definition. (W1 iv)

Q4. What is the practical example of ambiguity in labels? Please cite related work to support such a scenario. (W1 iii & v)

Q5. Narrow the scope in the title and the abstract to match the types of partial information treated in this work. The paper currently doesn't tackle partial information in its full generality or even the most practical cases. (W1 v)

Q6. Explain why the retain and forget risks are chosen to be cumulative not averaged. What would happen if they were averaged over the number of forget data points kk? (W1 vi)

Q7. Emphasise what type of perfect unlearning this work is targeting. Is it indistinguishability in the parameter space or the output space? (W2)

10 comments
AuthorsOpenReview ↗

Response to Reviewer 8mZ7 (Part 1/3)

We thank Reviewer 8mZ7 for careful reading and constructive feedback. We address each comment below.

Q1 (W1 iv): Definition and practicality of partial information

We acknowledge that our original description was insufficiently clear. The revised introduction explicitly defines F\mathcal F as a prescribed collection of observation functions F=f1,…,fm\mathcal F=\\{f_1,\ldots,f_m\\}, where each fj:D→Rf_j:\mathcal D\to\mathbb R, and explains how to extract information (such as original model, training samples, and Hessian) through the observation functions. We express the partial information as the tuple

IF(D):=(f1(D),…,fm(D)).\mathcal I_{\mathcal F}(D):=\bigl(f_1(D),\ldots,f_m(D)\bigr).

For example, as discussed in Footnote 1 in the revision, including the coordinate projections D↦(xi)jD\mapsto(x_i)_j for j∈1,…,dj\in\\{1,\ldots,d\\} and D↦yiD\mapsto y_i in F\mathcal F can encode the ii-th training sample in IF(D)\mathcal I_{\mathcal F}(D). Similarly, the original model w⋆(D)w^\star(D) trained on DD can be encoded by including fa(D):=[w⋆(D)]af_{a}(D):=[w^\star(D)]_{a} in F\mathcal F for every parameter index aa, and the Hessian at the original model w⋆(D)w^\star(D) can be encoded by including fab(D):=[∇w2L(w⋆(D);D)]abf_{ab}(D):=[\nabla_w^2 L(w^\star(D);D)]_{ab} in F\mathcal F for all parameter indices a,ba,b.

Q2 (W1 i, ii, iii, & v): What is and what is not the sufficient partial information for approximating the retrained model?

We thank the reviewer for this insightful question. First, we clarify that full access to the retained dataset is sufficient to recover the retrained reference. Specifically, the revised Lemma 2 in Section 3 (Lemma 3 in Appendix B.4 in the previous version) shows that if F\mathcal F contains the coordinate projections needed to recover every retained training sample, the unlearner can reconstruct the retained dataset and compute the retrained reference exactly. Consequently, both retain and forget risks are zero in our squared-loss setting.

Our analysis focuses on unlearning with incomplete access to the retained dataset. Such access restrictions can arise in practice. For example, in large language model unlearning, storing or retrieving the full training corpus at deletion time may be impractical. In federated unlearning, the server typically operates on client-provided model updates without directly accessing the raw training data held by individual clients. We also agree that using summaries for computational efficiency does not itself imply restricted data access, and we note that our analysis applies when the algorithm’s entire input is restricted to the specified observations.

Our construction highlights that even a small number of unavailable retained samples can prevent the unlearner from recovering the retrained reference. In Example 1, the differences between the hard instances are concentrated in kk retained samples (n−2k+1≤i≤n−kn-2k+1 \le i \le n-k), while the remaining n−2kn-2k retained samples (1≤i≤n−2k1 \le i \le n-2k) are identical. If these kk samples are unavailable, revealing all remaining retained data alongside the original model, the complete forget set, and the local Hessian still leaves the hard instances indistinguishable. The lower bounds in Theorem 1 therefore remain unchanged. Thus, even when the missing fraction kn−k\tfrac{k}{n-k} is arbitrarily small when n≫kn \gg k, the average forget-set prediction discrepancy remains bounded below by Rf/k≥25R_f/k\ge25, preventing recovery of the retrained reference in the worst case. We have discussed this point in the revised section 4.

More generally, directly characterizing minimal sufficient information is challenging because sufficiency depends on how tightly the available observations constrain the possible retraining targets for a desired approximation accuracy. This relationship varies with the admissible dataset class, model, and loss. Our analysis identifies concrete sufficient and insufficient information regimes, while a general characterization of minimal sufficient information remains an important direction for future work.

Q3 (W1 iv): Definition of information-extraction functions

We thank the reviewer for pointing out the confusion in our original description. As discussed in our response to Q1, in the revised introduction, we explicitly define F\mathcal F as a prescribed collection of observation functions F=f1,…,fm\mathcal F=\\{f_1,\ldots,f_m\\}, where each fj:D→Rf_j:\mathcal D\to\mathbb R. Thus, the Hessian at the original model w⋆(D)w^\star(D) can be encoded by including fab(D):=[∇w2L(w⋆(D);D)]abf_{ab}(D):=[\nabla_w^2 L(w^\star(D);D)]_{ab} in F\mathcal F for all parameter indices a,ba,b.

AuthorsOpenReview ↗

Response to Reviewer 8mZ7 (Part 2/3)

Q4 (W1 iii & v): Definition of label ambiguity

We thank the reviewer for the insightful question. We clarify that label ambiguity in our analysis concerns different datasets compatible with the same observations available to the unlearner, rather than uncertainty about labels in one fully observed training dataset. Each admissible dataset has fixed labels. At deletion time, however, the unlearner receives only IF(D)\mathcal I_{\mathcal F}(D), and multiple retained datasets may be consistent with this input. We note that restricted access at deletion time has been studied in the literature, e.g., Sekhari et al. [1] assume the unlearner uses the original model and stored statistics without the entire training dataset.

Section 3 illustrates this dataset ambiguity using two feasible datasets with identical retained features xix_i but different retained labels yiy_i. Because the unlearner receives the same inputs for both datasets, its output distribution must be the same. The risk decompositions quantify the resulting label variance and retrained-prediction variance across different feasible datasets.

Section 4 further establishes that such indistinguishable datasets can exist under the specified input restriction. Specifically, Example 1 constructs fully labeled feasible datasets with identical retained features xix_i but different retained labels yiy_i. These datasets share the same original model, forget set, and local Hessian, yet lead to different retraining targets. Thus, label variation provides a concrete way to demonstrate dataset ambiguity and establish worst-case lower bounds, without assuming that labels are ambiguous when the full labeled retained dataset is observed.

Q5 (W1 v): Clarification of the scope in the abstract

We thank the reviewer for the suggestion. We have revised the abstract to clarify that the partial information studied in this work refers to the limited knowledge of the training dataset available to the unlearning algorithm through the algorithm input, such as the original model, forgetting data, local Hessian of the training loss, and subset of retained data, but not the full retained data.

Q6 (W1 vi): Cumulative versus averaged risks

Following Guo et al. [2] and Izzo et al. [3], which use cumulative training losses, we define the retain and forget risks as cumulative losses over the retained and forget sets, respectively. Since nn and kk are fixed, the averaged risks are exactly

\overline R_f=\frac{R_f}{k}.$$ Theorem 1 therefore becomes $$\overline R_r\ge\frac{23}{8}\frac{k}{n-k}, \qquad \overline R_f\ge25,$$ and the retain upper bound becomes $\overline R_r\le Ck/(n-k)$. Thus the normalized forget bound is constant rather than growing with $k$, and the normalized retain bound depends on the deletion fraction. ### Q7 (W2): Approximating the retrained model in the parameter space versus the prediction space We acknowledge that our main analysis concerns prediction-space approximation, i.e., matching the retrained model’s empirical loss on the retained data and its predictions on the forget data. Such evaluation has precedents in the unlearning literature. For example, Golatkar et al. [4] measure $L_1$ differences between unlearned and retrained softmax outputs on both the retained and forget datasets, while Chundawat et al. [5] evaluate the average $L_2$ distance and JS divergence between unlearned and retrained predictions on the forget dataset. At the same time, in Appendix F we also establish a corresponding hardness result on parameter-space indistinguishability under the certified-removal framework of Guo et al. [2]. Specifically, Proposition 5 in Appendix F shows that, under the assumptions of deterministic retraining and a uniform bound $B$ on the squared prediction discrepancy per forget sample, any (randomized) unlearner satisfying $(\varepsilon,\delta)$-certified removal for every feasible dataset must satisfy $R_f(I_{\mathcal F})\le\delta kB$. Combining this with the lower bound $R_f(I_{\mathcal F})\ge25k$ from Example 1 yields $\delta\ge25/B$ for every $\varepsilon\ge0$. When $B$ is bounded independently of $n$ and $k$, this gives a constant lower bound on $\delta$, demonstrating that the information restriction also obstructs parameter-space indistinguishability under these assumptions. We highlighted this discussion in the revised section 3.
AuthorsOpenReview ↗

Response to Reviewer 8mZ7 (Part 3/3)

[1] Ayush Sekhari, Jayadev Acharya, Gautam Kamath, and Ananda Theertha Suresh. Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems, 2021.

[2] 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.

[3] Zachary Izzo, Mary Anne Smart, Kamalika Chaudhuri, and James Zou. Approximate data deletion from machine learning models. In International conference on artificial intelligence and statistics, pp. 2008–2016. PMLR, 2021.

[4] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Forgetting outside the box: Scrubbing deep networks of information accessible from input-output observations. In European Conference on Computer Vision, pp. 383–398. Springer, 2020.

[5] 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 AAAI Conference on Artificial Intelligence, pp. 7210–7217, 2023.

Reviewer_tc4sOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

the paper attempts to demonstrates that machine unlearning is fundamentally constrained by incomplete information and competing retention–forgetting objectives. Then it tried to provide bounds to characterize the performance of existing methods.

Critical strengths

the identified challenges for the existing unlearning literature is important and has been debate of the research such as Computational efficiency and incomplete data access.

Critical weaknesses

the paper is not properly organized. it is totally okay that the paper includes a short intro and and moving to the preliminaries and start the discussion about the proposed approach. However, this paper doesn't have any structure. From the intro they start introducing the IFI_F which is not a clear descrption or reference provided for it. Also they repeat the formulas from the intro in the prelimanries.

line 55 ) This description about the information is not correct. if the FF function is a classification function over the data, then it is not the PMF. for the self-information, the F needs to be a PMF function. The amount of information conveyed by an individual source xix_i with probability pip_i is known as its self-information or surprise, I(p_{i})}. But here if the classification model is considered therefore the self information can not be defined.

line 61) what does ambiguity mean? what is the feasible datasets?

line 68 ) the feasible dataset is not defined

line 135 ) what is partial information of the dataset?

Questions for authors

see previous section

1 comment
AuthorsOpenReview ↗

Response to Reviewer tc4s

We thank Reviewer tc4s for the detailed feedback. We have revised the introduction and preliminaries to improve clarity and address the comments below.

1. Organization of the paper

We thank the reviewer for the suggestion to improve the organization. In the revised introduction, we first explain what partial information is available to the unlearning algorithm, then formalize it through observation functions and introduce dataset ambiguity and feasible datasets. To reduce repetition, we have removed the repeated definitions of the dataset split and the original and retrained models from the preliminaries.

2. Definition of partial information

We acknowledge that our original description was insufficiently clear. Here, partial information refers to the limited knowledge of the training dataset available to the unlearning algorithm through the algorithm input, rather than Shannon self-information or surprisal. We represent this information through a prescribed collection of dataset-dependent observations.

The revised introduction explicitly defines the admissible dataset class D\mathcal D and a prescribed collection of observation functions F=f1,…,fm\mathcal F=\\{f_1,\ldots,f_m\\}, where each fj:D→Rf_j:\mathcal D\to\mathbb R. We express the available information as the tuple, which is also the input to the unlearning algorithm, defined as

IF(D):=(f1(D),…,fm(D)).\mathcal I_{\mathcal F}(D):=\bigl(f_1(D),\ldots,f_m(D)\bigr).

For example, as discussed in Footnote 1 in the revision, including the coordinate projections D↦(xi)jD\mapsto(x_i)_j for j∈1,…,dj\in\\{1,\ldots,d\\} and D↦yiD\mapsto y_i in F\mathcal F can encode the ii-th training sample in IF(D)\mathcal I_{\mathcal F}(D). Similarly, the original model w⋆(D)w^\star(D) trained on DD can be encoded by including fa(D):=[w⋆(D)]af_{a}(D):=[w^\star(D)]_{a} in F\mathcal F for every parameter index aa, and the Hessian at the original model w⋆(D)w^\star(D) can be encoded by including fab(D):=[∇w2L(w⋆(D);D)]abf_{ab}(D):=[\nabla_w^2 L(w^\star(D);D)]_{ab} in F\mathcal F for all parameter indices a,ba,b. In the main hardness construction, the available input to the unlearner consists of the original model, the complete forget set, and the local Hessian, while the full retained dataset is unavailable.

3. Definition of dataset ambiguity

In the revised introduction, we have clarified the definition of dataset ambiguity. For a restricted collection F\mathcal F, distinct datasets D,D′∈DD, D' \in \mathcal D may produce identical observations IF(D)=IF(D′)\mathcal I_{\mathcal F}(D) = \mathcal I_{\mathcal F}(D'). Dataset ambiguity refers to the inability to distinguish DD and D′D' using the observations available to the unlearning algorithm.

4. Definition of feasible datasets

In the revised introduction, we have also clarified the definition of feasible datasets. Given an observed input IFI_{\mathcal F}, the feasible datasets are all admissible datasets that could have produced this input:

S(IF):=D∈D:IF(D)=IF.S(I_{\mathcal F}):=\\{D\in\mathcal D:\mathcal I_{\mathcal F}(D)=I_{\mathcal F}\\}.

Because the unlearning algorithm receives the same input IFI_{\mathcal F} for every feasible dataset, it must induce the same output distribution. However, the retraining targets associated with different feasible datasets may differ, thus creating an obstruction to matching the retrained reference uniformly over S(IF)S(I_{\mathcal F}).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.