acceptodds
Withdrawn from ICLR 2027

Deep Repurposing: Reallocating Obsolete Representations Without Retraining

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Deep neural networks are increasingly deployed in long-lived systems, where task requirements may change after training. In such settings, part of the original output space may become obsolete: a class, prediction region, or learned behaviour may no longer be valid. Existing approaches either leave the obsolete behaviour intact or require fine-tuning, which can be expensive. We propose Deep Repurposing (DR), a post-hoc framework for adapting models under task obsolescence. DR estimates the latent geometry of obsolete and retained regions, removes obsolete-supporting components, and reallocates retained-compatible evidence through an analytic repair map without gradient updates. This yields repaired predictions and representations in which obsolete regions no longer act as valid outputs, while useful obsolete structure can support the retained task. Across multiple task settings, DR removes obsolete behaviour while preserving retained utility. More importantly, across classification benchmarks, DR matches or exceeds competing unlearning and editing baselines in retained accuracy, eliminates obsolete predictions, and adapts up to faster than competing unlearning methods.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Model Editing, Machine Unlearning, Latent Representations, Task Obsolescence

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_aRSbOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Deep Repurposing (DR), a post-hoc framework for adapting pretrained models when parts of their output space become obsolete, without retraining. However, I have significant concerns about both the paper’s presentation and its technical contributions, particularly the insufficient justification of its methodological novelty and the lack of convincing experimental validation of its core repurposing mechanism. I believe these issues are unlikely to be adequately addressed during the rebuttal phase.

Critical strengths
  1. The paper propose a practical solution for adapting models in settings where retraining is costly.
  2. The experiments cover a diverse range of datasets and tasks.
Critical weaknesses
  1. The Introduction and Related Work do not sufficiently establish the research gap or explain how the limitations of existing methods motivate the specific design of Deep Repurposing (DR). In particular, the central concept of “useful obsolete evidence” remains insufficiently defined, and the paper does not clearly explain (1) what constitutes such evidence, (2) how useful information shared between obsolete and retained regions can be identified and preserved, and (3) how DR fundamentally differs from existing feature-editing, decision-boundary modification, and output-remapping methods. Although Section 4 describes the proposed pipeline and the experiments provide some evidence of its effects, the paper does not fully establish the necessity or distinctiveness of its core repurposing mechanism.

  2. The experiments do not convincingly demonstrate that useful evidence is genuinely redirected to semantically appropriate retained regions. The ablations also do not sufficiently isolate the contribution of evidence redirection from obsolete-feature removal and residual refinement. Thus, the central claim of repurposing remains insufficiently validated.

Questions for authors
  1. For W1. While the Section 4 specifies an operational procedure for selecting a destination region, it remains unclear what makes the transferred obsolete evidence genuinely useful and how its semantic usefulness is independently validated.
  2. The discussion in Section 2 is largely descriptive and does not compare their mechanisms and limitations with those of DR. For example, what are the fundamental technical difference between [1] and DR? What distinct advantages does DR offer over other related methods [2-4]?
  3. Figure 4 provides a qualitative segmentation example, but quantitative results would strengthen the paper’s central claim of evidence repurposing, particularly that obsolete evidence is redirected toward semantically appropriate retained classes.

[1] Editing models with task arithmetic, ICLR 2023.

[2] Exploring and leveraging class vectors for classifier editing, NeurIPS 2025.

[3] Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary, CVPR 2023.

[4] Eternal sunshine of the spotless net: Selective forgetting in deep networks, CVPR 2020.

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

This paper studies how to adapt a deployed neural network when some output classes or prediction regions become obsolete. The proposed method, Deep Repurposing (DR), keeps pretrained weights frozen. It estimates latent prototypes and repair directions from calibration data, strips obsolete components, and shifts features toward retained regions. A residual correction is then applied before reconstructing the final predictions and latent vectors. The authors include two theoretical propositions and test DR on classification, regression, and segmentation. Across tasks, the method suppresses obsolete outputs at low calibration cost while largely preserving retained performance.

Critical strengths
  1. Relevant problem and lightweight setup. Removing retired or unreliable classes without retraining is a practical deployment problem. Freezing the pretrained model makes the method useful when full fine-tuning is too costly. The MSL and FishNet case studies in Appendix C give concrete examples of this setting.

  2. Broad experiments and useful ablations. The evaluation covers classification, regression, and segmentation rather than staying on simple image benchmarks. Beyond accuracy, the paper checks calibration, latent behavior, and runtime. Tables 15 and 16 in Appendix D are helpful because they separate residual refinement from the geometric repair.

Critical weaknesses
  1. Empirical gains are not clearly separated from residual correction and output suppression. On MSL (Table 15), retained accuracy stays at 0.803 without residual refinement and reaches 0.884 only with it. Logit suppression also matches DR on CIFAR-10 retained accuracy and F1 (Table 14), while obsolete suppression relies on the Eq. (23) score floor. The experiments also do not test semantic reassignment directly, as ADE20K (Table 13) uses unmerged labels.

  2. Theoretical guarantees do not cover the complete inference procedure. Proposition 1 only covers the intermediate vector zr+z^+_r, not the reconstructed representation z∗(x)z^*(x) in Eq. (26). Its proof assumes conditions (Eqs. 41 and 47) that Algorithm 2 never enforces at inference. Proposition 2 also guarantees obsolete suppression via the Eq. (23) score floor rather than proving geometric reallocation.

Questions for authors
  1. How much of the performance gain comes from geometric repurposing rather than residual refinement? Residual refinement still alters predictions when the repair mask M(x)M(x) is off, and Table 15 shows it drives the MSL gain. Could you compare DR against logit suppression with the same residual refinement to isolate the effect of geometric transport?

  2. Can the guarantees in Proposition 1 extend to the actual inference algorithm? Algorithm 2 does not enforce Eqs. (41) or (47), and Eq. (13) picks the minimum stripping step with no slack for transport. Does the margin bound still hold without those constraints, and does it carry over to z∗(x)z^*(x) in Eq. (26)?

  3. How do you verify that obsolete inputs are mapped to semantically valid retained classes? Table 4 only tracks nearest centroids, and ADE20K (Table 13) evaluates against the original unmerged labels. Could you evaluate ADE20K on the merged ground-truth taxonomy and compare against directly merging obsolete probabilities into target classes?

Reviewer_4cUwOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces Deep Repurposing (DR), a training-free, post-hoc framework for adapting pretrained neural networks when parts of their output space become obsolete. DR estimates the latent geometry of obsolete and retained regions using diagonal Gaussian prototypes, removes obsolete-supporting components via analytic stripping, and reallocates evidence toward compatible retained regions: all without gradient updates. The method is validated across classification (CIFAR10, CIFAR100, Flowers102, ImageNet), regression (UTKFace, Bike Sharing), and semantic segmentation (ADE20K), plus two real-world case studies (Mars rover, deep-sea fish). DR matches or exceeds unlearning and editing baselines on retained accuracy while eliminating obsolete predictions, and reports calibration times up to 60× faster than competing methods.

Critical strengths

Broad experimental coverage with strong retained accuracy. The evaluation spans nine datasets across four task types (classification, target-associated regression, semantic-region regression, segmentation) and multiple architectures. DR consistently preserves or improves retained accuracy while fully eliminating obsolete predictions; a combination no single baseline achieves across all settings.

Critical weaknesses

Critical Weaknesses

Missing required AI Use Statement. The ICLR 2027 Author Guidelines mandate an AI Use Statement for all submissions. This paper does not include one. The Ethics Statement and Reproducibility Statement, while only recommended, are also absent.

Imbalanced space allocation between method and evaluation. Sections 4.1 and 4.2 consume approximately 4 of the 9 allowed pages with dense mathematical formulations (~25 equations), leaving limited room for experimental evaluation. As a result, the experimental setup (Appendix F), ablation study (Appendix D), and two case studies (Appendix C) are all pushed to supplementary material. Reviewers should not need to rely on the appendix to assess core claims. The authors should compress the method derivation and reclaim space for a self-contained experimental section with setup, ablations, and critical case studies in the main body.

Incomplete and misleading runtime comparison. The "up to 60× faster" claim is featured in the Abstract but not clearly supported in the main body. The full timing comparison across all datasets and methods appears only in Table 7 in Appendix D. Beyond this, the comparison itself is not apples-to-apples: it measures DR's one-time calibration cost against finetuning/retraining one-time training costs, but omits DR's per-inference overhead. Unlike finetuning and retraining (which incur zero additional cost after training) DR runs latent stripping, contrastive transport, residual refinement, and SVD-based latent reconstruction at every input. No per-sample wall-clock inference time is reported. The claimed speedup therefore overstates the practical deployment advantage.


Minor Weaknesses

Segmentation showcase could be stronger. The ADE20K experiment in Figure 4 uses obsolete classes (armchair, swivel chair, stool, seat) that are subtypes of the retained parent chair: effectively a taxonomy-merging scenario rather than true obsolescence. The classification and regression experiments do test genuinely unrelated obsolete classes, so the method is not untested on harder cases. However, given the limited main-body space and the volume of material already in the appendix, featuring this easier task as the primary segmentation showcase in the core paper may not be the strongest editorial choice.

Individual components of DR from prior works. Deep Unlearning: Fast and Efficient Training-free Approach to Controlled Forgetting (Kodge, Saha & Roy, 2023) already covers post-hoc, training-free removal without retraining. Class Unlearning via Depth-Aware Removal of Forget-Specific Directions (Hatami, Aalishah & Monosov, 2026) is especially close, it addresses class-level latent direction removal with retention awareness. QR-Erase: Efficient Subspace-Based Machine Unlearning with Layer Localization (Lizzo & Heck, 2026) and POUR: A Provably Optimal Method for Unlearning Representations via Neural Collapse (Le et al., 2025) cover subspace/geometric projection-based forgetting. RePAIR: Interactive Machine Unlearning through Prompt-Aware Model Repair (Rachapudi et al., 2026) and SAGE: Retain-Aware Post-Hoc Sanitization of Final Unlearning Vector (Zhang et al., 2026) cover post-hoc analytic repair with retain-aware correction. I suggest to clarify formal distinction, the contribution may read as a combination of known techniques rather than a new method.

Questions for authors

Evidence floor robustness (Eq 23). The obsolete evidence floor is defined as e*_o = 2·min_r(e*_r) − log(1+|R|ε). For this to guarantee suppression below all retained scores, it requires min_r(e*_r) < 0. Since e*_r = log(q*_r + ε) and q*_r ∈ [0,1], this typically holds, but when a single retained class dominates (q*_r → 1), the floor can approach zero or become positive. Proposition 2 derives this under the simplex constraint but does not discuss this edge case. Can the authors clarify under what conditions, if any, the floor fails to suppress obsolete scores?

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

Related papers

Discussion (0)

Sign in to comment.