acceptodds
Under review as a conference paper at ICLR 2027

Spectral Saliency for Machine Unlearning

Abstract

Machine unlearning (MU) aims to remove the influence of specific training data while preserving model utility. As the name suggests, MU can be viewed as the inverse of learning, using gradient-based updates to reduce the influence of a forget-set by counteracting the previously learned behavior. Recently, Muon, a gradient descent variant, has been introduced. Muon applies spectral magnitude normalization to encourage exploration of rare directions and demonstrates promising performance. Inspired by Muon, we adopt the spectral view for unlearning and propose Spectral Saliency Unlearning (SSU). SSU thresholds weak singular components and updates only those directions supported by a confident unlearning signal. We further provide theoretical justification for this thresholding approach from the perspective of the forgetting-retention trade-off. Experiments across image classifiers, diffusion models, and LLMs demonstrate SSU's effectiveness.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Machine unlearning, spectral saliency, optimization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_6tArOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a fine-grained, gradient-based unlearning that utilizes SVD for the unlearning objective gradient refinement.

The proposed method uses SVD to decompose a combined unlearning objective gradient (Gu=Gf+GrG_u = G_f + G_r), and discards components correspond to small singular values. The hypothesizes is that those components are directions that have gradient conflict between the forget and retain objectives.

The paper then provides substantial amount of derivations to support their SVD-based gradient filtering.

The paper also provides experimental evidence over a diverse set of models: image classifiers, image diffusion models, and LLMs.

Critical strengths
  • The method is simple and concise. The proposed algorithm recipe is easy to follow and understand: main thing is to apply SVD decomposition over the computed unlearning gradients, and tune a singular value threshold for gradient refinement.

  • Experiment covers models for different tasks. The proposed method is demonstrated work for image classifier, image diffusion models, and LLMs.

Critical weaknesses
  • Lack of significant technical and engineering details. The paper delivers method and derivations utilizing unlearning objective Lu\mathcal{L}_u, which breaks down to Lu=Lf+Lr\mathcal{L}_u = \mathcal{L}_f+\mathcal{L}_r, but these objectives is never concretely written or defined.

    • For example, is Lr\mathcal{L}_r and Lf\mathcal{L}_f actually in the same form but just the sign being different, or they are identical just the input data D\mathcal{D} being different?
    • This is important because the paper later on experiments their method on models of different tasks, and each of them has a significantly different training objectives.
    • In Algorithm 1, the superscript ℓ\ell appeared in line-4 but never introduced anywhere else. Does it correspond to layer of the networks in the actual implementation? If so, does the singular value threshold also needs to be tuned as ℓ\ell-wise (i.e., different G(ℓ)G^{(\ell)} will need a different kk), then how many layers did they consider? In this case the proposed method appeared to be too heuristic, which looks nice at high-level notations, but requires heavy engineering effort when it comes to the actual implementation (apparently it doesn't scale if you think of a large model that have many layers --- it is also non-transparent how you did that on LLMs).
  • Lack of basic experimental evidence. The key assumption the paper makes is that the singular-value structure of the sum of two gradients reveals the alignment or conflict between the individual gradient (GfG_f and GrG_r), yet the paper shows no direct evidence on that. Even though the paper currently provided pack of derivations and experiments, but it will be more convincing if the paper can have direct evidence from basic gradient analysis:

    • Measure the forget and retain graidents separately in the combined SVD basis: for each Di=uivi⊤D_i = u_iv_i^\top, report <Gf,Di><G_f, D_i> and <Gr,Di><G_r, D_i>, check if <Gu,Di>=<Gf,Di>+<Gr,Di><G_u, D_i> = <G_f, D_i> + <G_r, D_i> holds at first hand. Then test whether low-singular-value components actually show more opposition than high-singular-value components.
    • Back to the first bullet, the paper never defines Lf\mathcal{L}_f and Lr\mathcal{L}_r so it is also unclear how they can do such a gradient check.

[Minor]

  • In Line 88-89, the notation for two gradient should specified what variable that the gradient is with respect to (reviewer assumes it is ∇WL\nabla_W\mathcal{L})
Questions for authors
  • Could you provide technical details, regarding how are Lr\mathcal{L}_r and Lf\mathcal{L}_f defined?

    • Subquestion: how are Lr\mathcal{L}_r and Lf\mathcal{L}_f defined for each task model (image classifiers, image diffusions, LLMs)
  • Could you provide sufficient implementation details for better understanding of the experiment and its reproducibility, beyond the current Appendix B?

  • Could you provide basic gradient analysis and its implementation details, as specified in the second bullet of the Critical Weaknesses?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 6tAr (1/3)

We sincerely thank Reviewer 6tAr for their feedback. We answer W1 (with Q1, Q2 and the minor point) in our first two comments and W2 (with Q3) in the third comment, and sincerely hope the reviewer will consider raising the score if our reply help address some of their concerns.

1. Definitions of Lf\mathcal L_f and Lr\mathcal L_r (W1, Q1). We respectfully clarify that in every SSU experiment both terms are minimized, with no sign flip; they differ in data and targets and, for the diffusion models and LLMs, also in form. The objective is Lu=cfLf+crLr\mathcal L_u=c_f\mathcal L_f+c_r\mathcal L_r, where the weights cfc_f and crc_r follow existing works: SalUn [1] for classification, DDPM and Stable Diffusion, and Yuan et al. [2] for TOFU. The two terms are defined as follows (sg: stop-gradient; xtx_t: noised input at step tt):

  • Classification (SalUn's objective): Lf=CE(fW(x),y~)\mathcal L_f=\mathrm{CE}(f_W(x),\tilde y) on forget images with a random label y~\tilde y, and Lr=CE(fW(x),y)\mathcal L_r=\mathrm{CE}(f_W(x),y) on retain images: same form, different data and targets.
  • DDPM: Lr=∥ϵ−ϵW(xt,t,y)∥2\mathcal L_r=\Vert\epsilon-\epsilon_W(x_t,t,y)\Vert^2 on retained classes, and Lf=∥ϵW(xt,t,yf)−sg[ϵW(xt,t,y~)]∥2\mathcal L_f=\Vert\epsilon_W(x_t,t,y_f)-\mathrm{sg}[\epsilon_W(x_t,t,\tilde y)]\Vert^2 on images of the forgotten class yfy_f, with a fixed other class y~\tilde y.
  • Stable Diffusion: the same pair in latent space, conditioned on CLIP embeddings of class prompts.
  • TOFU, ME+GD: Lf\mathcal L_f is the token mean of KL(U ∥ pW(⋅∣prefix))\mathrm{KL}(\mathcal U\,\Vert\,p_W(\cdot\mid\text{prefix})) on forget samples, with U\mathcal U uniform over the vocabulary; Lr\mathcal L_r is the token-mean NLL of retain answers.
  • TOFU, IDK+AP: Lf\mathcal L_f is the token-mean NLL of an "I don't know" answer to a forget question, and Lr=−1βElog⁡σ(β[log⁡pW(a∣q)−log⁡pW(aidk∣q)])\mathcal L_r=-\frac1\beta\mathbb E\log\sigma(\beta[\log p_W(a\mid q)-\log p_W(a_{\rm idk}\mid q)]) on retain questions.

Both losses are explicit, so GfG_f and GrG_r can be computed separately per weight matrix at any iterate, as in our response 5. We will clarify these definitions, including that both terms are minimized in every task, in the revised version.

2. The index ℓ\ell, thresholds and number of layers (W1). ℓ\ell indexes the weight matrices (layers). G(ℓ)=∇W(ℓ)LuG^{(\ell)}=\nabla_{W^{(\ell)}}\mathcal L_u is reshaped to mℓ×nℓm_\ell\times n_\ell (a convolution kernel to out × in·kHkWk_Hk_W) and keeps kℓ=max⁡(1,⌊γmin⁡(mℓ,nℓ)⌋)k_\ell=\max(1,\lfloor\gamma\min(m_\ell,n_\ell)\rfloor) components. We respectfully clarify that γ\gamma is a single fixed number per run, shared by all layers: nothing is tuned per layer, and kℓk_\ell varies only with the matrix size. The number of hyperparameters therefore does not grow with depth. We will define ℓ\ell, kℓk_\ell and the shared γ\gamma in Algorithm 1.

[1] Fan et al., SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation, ICLR 2024, arXiv:2310.12508.

[2] Yuan et al., A Closer Look at Machine Unlearning for Large Language Models, ICLR 2025, arXiv:2410.08109.

AuthorsOpenReview ↗

Response to Reviewer 6tAr (2/3)

3. Actual implementation and layers considered (W1, Q2). For each task model, the layers we consider are fixed by layer type:

  • Classification (ResNet-18): SSU filters the gradients of all 20 convolutions and the final linear layer with an exact SVD at every step. For the BatchNorm and bias parameters, which are vectors, SSU instead keeps the fraction γ\gamma of gradient coordinates with the largest magnitude.
  • DDPM: SSU filters the 89 convolution kernels of the U-Net with an exact SVD at every step, and all other parameters use a coordinate mask computed once from the forget-set gradient.
  • Stable Diffusion: SSU filters the 64 cross-attention matrices. Following SalUn, we compute the saliency scores once before unlearning as the entrywise magnitude of the forget-set gradient. At every step, we then project the gradient of each matrix onto the top singular directions of its saliency scores. All other parameters use a coordinate mask computed from the same scores.
  • Llama-2-7B: SSU filters the 96 q/k/v projections of the 32 decoder layers with a randomized rank-kk SVD on every fifth micro-batch. All other parameters use a coordinate mask computed once from the forget-set gradient, which is also applied to q/k/v on the other micro-batches.

o address the concern that SSU may not scale to a large model with many layers, we timed the SVD routine of our training code for the full Llama-2-7B model on one A10G (median of repeated calls after warm-up):

SVD of q/k/v (4096 × 4096)Time ↓Peak extra GPU memory ↓
Exact, one matrix3.86 s386 MiB
Randomized rank-819, one matrix62.6 ms333 MiB
Randomized, all 96 matrices6.0 s333 MiB
Amortized over five micro-batches1.2 s per micro-batch–

We observe that the randomized SVD is about 60× faster than an exact one, and since the matrices are filtered one at a time, the peak memory grows only by one matrix's workspace. We will include these discussion in Sec. 5.3 of the revised version.

Finally, for every task model the filtered gradient is passed to the optimizer of the base method (SGD with momentum, Adam or AdamW), and the plain gradient step in Algorithm 1 is a simplification for presentation. We will state this in Algorithm 1 and extend Appendix B with the matrix shapes, kℓk_\ell, step structure and optimizer settings of every task model. We will release the code upon acceptance.

4. Gradient notation (Minor). We sincerely thank the reviewer for pointing this out. The gradients are taken with respect to the weights, Gf=∇WLf(W;Df)G_f=\nabla_W\mathcal L_f(W;\mathcal D_f); in practice Gf(ℓ)=∇W(ℓ)Lf(Wt;Bf)G_f^{(\ell)}=\nabla_{W^{(\ell)}}\mathcal L_f(W_t;\mathcal B_f) per weight matrix on a mini-batch (likewise GrG_r and GuG_u). We will correct this in the updated revision.

AuthorsOpenReview ↗

Response to Reviewer 6tAr (3/3)

5. Basic gradient analysis (W2, Q3). We thank the reviewer for this suggestion and have run the requested analysis. Specifically, on Tiny-ImageNet (three seeds) and DDPM (SSU on the convolution kernels or on all layers), we project the forget and retain gradients onto each singular direction Di=uivi⊤D_i=u_iv_i^\top of the combined gradient Gu=Gf+Gr=∑iσiuivi⊤G_u=G_f+G_r=\sum_i\sigma_iu_iv_i^\top, i.e., fi=⟨Gf,Di⟩f_i=\langle G_f,D_i\rangle and ri=⟨Gr,Di⟩r_i=\langle G_r,D_i\rangle, and first verify that ⟨Gu,Di⟩=fi+ri\langle G_u,D_i\rangle=f_i+r_i holds (up to floating-point error). We then compare the directions SSU keeps with those it truncates, before and after unlearning, by the share of opposed directions (firi<0f_ir_i<0), the mean alignment a(vi)a(v_i) of Def. 4.1, and the first-order progress on each loss (∑σifi\sum\sigma_if_i and ∑σiri\sum\sigma_ir_i, in units of ∥Gu∥2\Vert G_u\Vert^2). The following table reports kept / truncated values; parentheses count the cases in which the truncated directions are more opposed:

SettingOpposed share firi<0f_ir_i<0Alignment a(vi)a(v_i) (Def. 4.1)Progress on Lf\mathcal L_fProgress on Lr\mathcal L_r
Tiny-ImageNet, start0.051 / 0.056 (3/3)0.117 / −0.0110.936 / 0.05610.007 / 0.0009
Tiny-ImageNet, end0.000 / 0.007 (3/3)0.046 / −0.2930.398 / 0.02780.526 / 0.0475
DDPM conv-only, start0.025 / 0.061 (3/3)−0.085 / −0.4310.102 / 0.01240.759 / 0.1267
DDPM conv-only, end0.120 / 0.124 (7/9)−0.033 / −0.3260.004 / 0.00050.863 / 0.1331
DDPM all-layers, start0.025 / 0.058 (3/3)−0.173 / −0.4590.105 / 0.01220.758 / 0.1243
DDPM all-layers, end0.102 / 0.115 (7/7)−0.116 / −0.3520.004 / 0.00050.864 / 0.1313

We observe that:

  • The truncated directions are the least aligned. a(vi)a(v_i) is lower for the truncated directions in every case at every probe point of both pipelines, as predicted by Prop. 4.3. On Tiny-ImageNet they are also more often opposed on every seed throughout. On DDPM they are more often opposed before unlearning; during unlearning, opposition grows in every σ\sigma decile, most in the top one, so this sign test is mixed from step 250 on (end-point counts in the table).
  • Truncation costs little progress. The kept directions carry 86–94% of the first-order progress on each loss; the truncated ones hold only 6–8% (Tiny-ImageNet) or 13–15% (DDPM) of ∥Gu∥2\Vert G_u\Vert^2.
  • The leading directions carry the unlearning update. On Tiny-ImageNet, keeping as many directions from the bottom of the spectrum instead of the top raises the Avg. Gap from 5.35 to 26.49.

Implementation details of the analysis (Q3). We run the analysis on Tiny-ImageNet and DDPM, because in these two pipelines each SSU step filters one combined gradient, which is the setting analysed in Sec. 4. On Tiny-ImageNet, forget and retain images share each mini-batch, and on DDPM, each step uses one joint loss. We measure the gradients before unlearning, at six points along the unlearning run, and on the final checkpoints of the reported runs. All measurements use fixed probe batches: four mixed batches of 256 images per seed on Tiny-ImageNet, and eight pairs of 128 forget and 128 retained images per class on DDPM, with fixed noise and timesteps. For every filtered matrix, we compute GfG_f, GrG_r and GuG_u with three backward passes through the same graph, using the losses and weights of the unlearning step. Since GuG_u is then the gradient that SSU filters at that step, the kept directions are exactly those SSU keeps.

We thank the reviewer for their valuable feedback and will incorporate these clarifications into the revision.

Reviewer_ZesnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • This paper investigates an optimization during unlearning. The paper utilizes SVD to decompose the components in spectral view and proposes the Spectral Saliency Unlearning method. The proposed method is a plug-and-play method on existing unlearning methods. The experimental results show the effectiveness on various data and domains, e.g., image classification, generation, and LLMs.
Critical strengths
  • The paper investigates an underestimated point of unlearning about optimization.
  • The paper got strong motivation from existing optimization techniques, such as SignSGD and Muon.
  • The paper contains the theoretical verification of the proposed method.
Critical weaknesses
  • Please refer to the Questions below.
Questions for authors
  • From my perspective, the proposed SSU, which is a plug-and-play method after backpropagation and before gradient update, is a simple version of truncated SVD rather than unlearning variants from sign-SGD or Muon. This motivation is overstated.
  • To ensure the effectiveness of the proposed method, showing each alignment or training dynamics of truncated and non-truncated components w.r.t. forget-retain gradients is important to back up the theoretical justification in Section 4.
  • As stated in the Background section, the recent optimizer Muon does not truncate small spectral components, rather matching the magnitude of each component by not using the Σ\Sigma. If it is the case for “learning”, then why does the proposed method have a contradictory idea based on my understanding?
  • Following the previous question, did the authors perform the analysis of Σ\Sigma for unlearning? Can the authors explain the success/failure of the method based on the understanding of this paper? See [1] for Muon+Unlearning.
  • In SVD, finding the elbow, which explains the ratio that the principal components explain, is very important. In Table 7 or Appendix B, it seems like the selection of γ\gamma varies in different settings.
  • The results in the experimental sections show the effectiveness of the proposed method on top of the strongest baseline method. Is it applicable to other baselines as well?
  • The computation of SVD is heavy for large models. Although the paper reports limited overhead for image models, what are the additional runtime and memory costs for LLMs?

[1] Subspace Control: Turning Constrained Model Steering into Controllable Spectral Optimization, COLM 2026.

5 comments
AuthorsOpenReview ↗

Response to Reviewer Zesn (1/4)

We thank the reviewer for their comments. We answer Q1–Q7 in order, and sincerely hope the reviewer will consider raising the score if our reply addresses some of your concerns.

Q1. Motivation. We thank the reviewer for noting our motivation from SignSGD and Muon as a strength. We agree that, as an operation, SSU is a rank-kk truncated SVD of each layer's gradient (Algorithm 1), and we did not intend it as a variant of SignSGD or Muon. Muon is what motivated it: it showed that acting on the singular values of the layer gradient, rather than on its coordinates, is effective for training. SSU carries this spectral view to unlearning: SignSGD equalizes and SalUn thresholds coordinate magnitudes; Muon equalizes and SSU thresholds singular values. Q3 shows why unlearning calls for thresholding, and that the two are complementary. We will make this relationship explicit in the introduction.

Q2. Alignment and training dynamics. We thank the reviewer for this suggestion and agree that this analysis directly supports Sec. 4, so we have added it. Specifically, on Tiny-ImageNet (three seeds) and DDPM (SSU on the convolution kernels or on all layers), we project the forget and retain gradients onto each singular direction Di=uivi⊤D_i=u_iv_i^\top of the combined gradient Gu=Gf+Gr=∑iσiuivi⊤G_u=G_f+G_r=\sum_i\sigma_iu_iv_i^\top, i.e., fi=⟨Gf,Di⟩f_i=\langle G_f,D_i\rangle and ri=⟨Gr,Di⟩r_i=\langle G_r,D_i\rangle, and compare the directions SSU keeps with those it truncates, before and after unlearning, by the share of opposed directions (firi<0f_ir_i<0), the mean alignment a(vi)a(v_i) of Def. 4.1, and the first-order progress on each loss (∑σifi\sum\sigma_if_i and ∑σiri\sum\sigma_ir_i, in units of ∥Gu∥2\Vert G_u\Vert^2). The following table reports kept / truncated values; parentheses count the cases in which the truncated directions are more opposed:

SettingOpposed share firi<0f_ir_i<0Alignment a(vi)a(v_i) (Def. 4.1)Progress on Lf\mathcal L_fProgress on Lr\mathcal L_r
Tiny-ImageNet, start0.051 / 0.056 (3/3)0.117 / −0.0110.936 / 0.05610.007 / 0.0009
Tiny-ImageNet, end0.000 / 0.007 (3/3)0.046 / −0.2930.398 / 0.02780.526 / 0.0475
DDPM conv-only, start0.025 / 0.061 (3/3)−0.085 / −0.4310.102 / 0.01240.759 / 0.1267
DDPM conv-only, end0.120 / 0.124 (7/9)−0.033 / −0.3260.004 / 0.00050.863 / 0.1331
DDPM all-layers, start0.025 / 0.058 (3/3)−0.173 / −0.4590.105 / 0.01220.758 / 0.1243
DDPM all-layers, end0.102 / 0.115 (7/7)−0.116 / −0.3520.004 / 0.00050.864 / 0.1313

We observe that:

  • The truncated directions are the least aligned. a(vi)a(v_i) is lower for the truncated directions in every case at every probe point of both pipelines, as predicted by Prop. 4.3. On Tiny-ImageNet they are also more often opposed on every seed throughout. On DDPM they are more often opposed before unlearning; during unlearning, opposition grows in every σ\sigma decile, most in the top one, so this sign test is mixed from step 250 on (end-point counts in the table).
  • Truncation costs little progress. The kept directions carry 86–94% of the first-order progress on each loss; the truncated ones hold only 6–8% (Tiny-ImageNet) or 13–15% (DDPM) of ∥Gu∥2\Vert G_u\Vert^2.
  • The leading directions carry the unlearning update. On Tiny-ImageNet, keeping as many directions from the bottom of the spectrum instead of the top raises the Avg. Gap from 5.35 to 26.49 (three seeds; Q5).

We will add the results in the revised paper.

AuthorsOpenReview ↗

Response to Reviewer Zesn (2/4)

Q3. Learning vs. unlearning. We would bring to the reviewer's attention that the two settings differ in the number of objectives. Muon learns a single objective, where weak singular directions carry weaker, more specific structure that equalization promotes (Sec. 4.1). Unlearning minimizes two competing losses, and Prop. 4.3(a) shows that a singular value of Gu=Gf+GrG_u=G_f+G_r that is small relative to the forget/retain signal along its direction implies negative forget–retain alignment; equalizing would raise exactly these directions to the magnitude of the leading ones. To test this, we ran equalization controls on Tiny-ImageNet (γ=0.5\gamma=0.5, three seeds). An equalized update sets all its singular values to their root mean square, which keeps its Frobenius norm; every filtered cell keeps the top 50% of the one-dimensional gradients, as SSU does, while RL filters nothing:

Avg. Gap ↓all directionstop-kk directions
raw singular values6.57 ± 0.62 (RL)5.35 ± 0.22 (SSU)
equalized singular values3.84 ± 0.363.20 ± 0.85

We observe that truncating the weak directions lowers the Avg. Gap both without and with equalization. The best update truncates and then equalizes the kept directions, which is also the structure of the Muon-based unlearning method of [1] (Q4), so truncation and equalization are complementary in unlearning. We will add these results to Sec. 5.4.

Q4. Analysis of Σ\Sigma, success/failure, and relation to [1].

Q4.1: Did we analyze Σ\Sigma for unlearning? We computed the singular value spectra of GuG_u, GfG_f and GrG_r for every weight matrix at the probe points of Q2. The spectra are steep: the elbow of GuG_u lies at 14–21% of the rank, already holds 82–90% of the energy ∑iσi2\sum_i\sigma_i^2, and barely moves during unlearning. With γ=0.5\gamma=0.5, SSU therefore keeps 96–99% of the energy of a typical matrix and removes only its flat tail.

Q4.2: When does SSU succeed or fail? Our γ\gamma sweep (Q5) explains this through the spectrum. SSU helps as long as it keeps all directions up to the elbow and removes only the flat tail: γ=0.3\gamma=0.3 and 0.50.5 lower the Tiny-ImageNet Avg. Gap of the unfiltered update on every seed (6.57 → 4.54 and 5.35), mainly by moving FA and MIA toward Retrain while RA and TA stay nearly unchanged. Cutting below the elbow (γ=0.1\gamma=0.1) removes about 40% of the update's energy and hurts, while keeping nearly the whole spectrum (γ≥0.7\gamma\ge0.7) leaves the update close to unfiltered and gives no gain.

Q4.3: Relation to [1]. SIFT [1] is a Muon-based optimizer that is also applied to LLM unlearning. Where the gradients of the two objectives conflict, it restricts its update to the top-KK singular subspaces of their momenta (K=128K=128 for 4096-dimensional matrices), and it reports that unlearning is best at K=128K=128 and degrades for K>256K>256. [1] thus also finds that unlearning works best when weak spectral components are left out of the update, consistent with SSU. Unlike [1], SSU needs only the spectrum of the combined gradient, without separate gradients, an alignment test or Newton–Schulz iterations, and it keeps the magnitudes of the kept directions. We will discuss [1] in Related Work and add the spectra analysis in he revised version.

[1] Huang et al., Subspace Control: Turning Constrained Model Steering into Controllable Spectral Optimization, COLM 2026, arXiv:2604.04231.

AuthorsOpenReview ↗

Response to Reviewer Zesn (3/4)

Q5. Elbow and the choice of γ\gamma. γ\gamma is one fixed number per run, shared by all layers: a matrix with rℓr_\ell singular values keeps kℓ=max⁡(1,⌊γrℓ⌋)k_\ell=\max(1,\lfloor\gamma r_\ell\rfloor) components. The main results use γ=0.7\gamma=0.7 on CIFAR-10, 0.5 on Tiny-ImageNet, DDPM and Stable Diffusion, and 0.2 for the Llama-2 q/k/v projections. To relate γ\gamma to the elbow, we ran a more exhaustive search over γ\gamma on Tiny-ImageNet: the γ=0.5\gamma=0.5 SSU setting of Table 3 with only γ\gamma changed, three seeds each (γ=1\gamma=1 is the unfiltered RL update). We also kept as many directions as γ=0.5\gamma=0.5 but chose the bottom ones or a random subset. In every filtered row, the one-dimensional gradients keep their largest γ\gamma fraction of entries, as in SSU.

Directions kept per matrixAvg. Gap ↓
top 10% (γ=0.1\gamma=0.1)8.49 ± 0.95
top 30% (γ=0.3\gamma=0.3)4.54 ± 0.52
top 50% (γ=0.5\gamma=0.5, Table 3)5.35 ± 0.22
top 70% (γ=0.7\gamma=0.7)6.58 ± 0.48
top 90% (γ=0.9\gamma=0.9)7.13 ± 0.50
all (γ=1\gamma=1, RL)6.57 ± 0.62
bottom 50%26.49 ± 0.74
random 50%6.76 ± 0.96
  • Elbow. The median elbow of GuG_u lies at 20.3% of the rank on Tiny-ImageNet (Q4). γ=0.3\gamma=0.3 is the smallest swept value that keeps every direction up to it, and the search finds that it lowers the gap below both RL and our reported γ=0.5\gamma=0.5 on all three seeds. γ=0.1\gamma=0.1 cuts below the elbow and is worse than no filter on all three seeds. At the extreme end, γ=0.7\gamma=0.7 and 0.9 keep nearly the whole spectrum (≥98.6% of a median matrix's ∑iσi2\sum_i\sigma_i^2), weak directions included, and give no gain over the unfiltered update.
  • Ordering. With the same number of directions, keeping the bottom half gives 26.49, a random half 6.76, close to RL, and the top half 5.35. The bottom half carries only 3–4% of a median matrix's ∑iσi2\sum_i\sigma_i^2, so the bottom-kk update is small: the leading directions carry the unlearning update. Which directions are removed thus matters more than how many: removing the weakest half lowers the gap from 6.57 to 5.35, while removing a random half does not lower it on average.

We will add the sweep and this discussion to Sec. 5.4.

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

The paper proposes Spectral Saliency Unlearning (SSU), a plug-in change for gradient-based machine unlearning. At each step, it takes the singular value decomposition of each layer's combined unlearning gradient. It keeps only the top singular components and updates with this truncated gradient.

Validation covers:

  1. Image classification: CIFAR-10 (10% and 50% forgotten) and Tiny-ImageNet; SSU is added on top of SalUn
  2. Image generation: DDPM on CIFAR-10 and Stable Diffusion on Imagenette; SSU is added on top of SalUn
  3. LLMs: TOFU benchmark with Llama2-7B; SSU is added on top of ME+GD and IDK+AP

There are also ablations on the keep ratio and on which layers SVD is applied to.

Critical strengths

Simple and "cheap" to include plug-in. SSU adds a single hyperparameter, needs no extra data passes or offline saliency computation, and is evaluated on classifiers, two diffusion models and a 7B language model.

Critical weaknesses

None

Questions for authors
  1. What is the performance of SSU on a more complex image generation benchmarks like Style removals and NSFW?
  2. Validation of SSU's gains is limited to SalUn for image classification and image generation, so the claimed generality as a plug-in is not confirmed. Could you add at least 2 more gains checks per task?
1 comment

We sincerely thank the reviewer for the their comments. We address the two questions below with new gain checks of SSU on two additional base methods for both classification and generation, and sincerely hope the reviewer will consider raising the score.

1. Gains on other base methods (Q2). We added SSU to two more base methods per task: random labeling (RL, the objective of SalUn [1] without its mask) and SFRON [2]. The base method is unchanged (in terms of e.g., loss, data, schedule, hyperparameters). SSU replaces the gradient of every convolution kernel by its top-kk SVD reconstruction at every step, and all other parameters are treated as in the base method. Each method uses three seeds.

Image. We test on classification on Tiny-ImageNet using ResNet-18 with 10% random forgetting. We also include SalUn itself, with SSU filtering the convolution kernels on top of SalUn's mask:

MethodFARATAMIAAvg. Gap ↓
Retrain59.42 ± 0.5199.98 ± 0.0059.90 ± 0.0473.97 ± 0.690.00
SalUn74.08 ± 3.5799.76 ± 0.0953.30 ± 0.3482.06 ± 1.717.40 ± 0.57
SalUn + SSU57.50 ± 2.9299.49 ± 0.0452.35 ± 0.3386.24 ± 1.555.57 ± 0.80
RL53.20 ± 1.3899.77 ± 0.0351.34 ± 0.9585.25 ± 3.176.57 ± 0.62
RL + SSU55.23 ± 0.9399.71 ± 0.0550.94 ± 0.8583.79 ± 2.825.81 ± 0.06
SFRON98.83 ± 0.3399.98 ± 0.0060.67 ± 0.2223.50 ± 4.1122.66 ± 0.83
SFRON + SSU95.95 ± 0.5399.98 ± 0.0060.35 ± 0.2540.30 ± 4.2317.66 ± 0.96

We observe that SSU lowers the Avg. Gap of all three base methods on all three seeds. The γ=0.5\gamma=0.5 SSU row of Table 3, which uses the same RL objective and filters all 21 weight matrices, reaches 5.35 ± 0.22.

LLMs. For LLMs, Table 6 in the paper already applies SSU to two base methods (ME+GD and IDK+AP).

We can see that SSU consistently improves the performance across different base methods, supporting its generality as a plug-in.

2. Style removal and NSFW (Q1). We thank the reviewer for this suggestion. Since SSU only filters the gradient of the base method's objective, the same design applies to the objective of these tasks. However, each benchmark needs its own Stable Diffusion pipeline and large-scale generation for evaluation (NSFW: one image per I2P prompt, 4,703 prompts, for each model), and style removal additionally needs a style-fine-tuned model with its style classifiers. We could not complete these within the rebuttal window and prioritized Q2, which tests the plug-in claim on two additional base methods. We will add NSFW and style-removal results to the revision.

We thank the reviewer for their valuable feedback and will incorporate these results into the revision.

[1] Fan et al., SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation, ICLR 2024, arXiv:2310.12508.

[2] Huang et al., Unified Gradient-Based Machine Unlearning with Remain Geometry Enhancement, NeurIPS 2024, arXiv:2409.19732.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.