PaperlayerView exact revision on arXiv

Source-linked AI summary

Repurposing Obsolete Representations for Post-Deployment Adaptation

Daniel Bethell, Charmaine Barker, Simos Gerasimou

arXiv:2610.01453v1cs.LG

TL;DR

When deployment changes make part of a task obsolete, existing approaches may leave obsolete outputs in place or require costly retraining. DR uses calibrated latent geometry to remove obsolete support and redirect useful evidence to retained regions without updating model parameters. Across multiple task settings, it suppresses obsolete behaviour while preserving retained performance, and classification results show retained performance preserved or improved at low repair cost.

  • Problem

    Task obsolescence makes a previously valid class, prediction region, or learned behaviour invalid or no longer actionable as deployment conditions change.

    Source: §1 · p. 1

  • Method

    DR estimates region geometry and residual corrections from calibration data, then analytically repairs predictions and representations using fixed model parameters.

    Source: §3 · p. 2, §4.1 · p. 4, §4.2 · p. 6

  • Results

    Across multiple task settings, DR reduces obsolete behaviour while preserving retained performance; across classification datasets it preserves or improves retained performance while suppressing obsolete behaviour at low repair cost.

    Source: §1 · p. 2, §5.1 · p. 8

  • Takeaways & Limitations

    The results indicate that obsolete structure can be redirected toward retained tasks rather than simply masked, forgotten, or retrained away.

    Source: §6 · p. 9

    The setup assumes supervised train, calibration, and test data and keeps pretrained parameters fixed during repair and deployment adaptation.

    Source: §3 · p. 2

Abstract

from arXiv · show

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 $60\times$ faster than competing unlearning methods.

1 INTRODUCTION

Task obsolescence occurs when an output class, prediction region, or learned behaviour no longer fits the deployment objective. Deep Repurposing (DR) addresses this by removing obsolete identity while reallocating useful evidence to retained regions.

1 INTRODUCTION

  • A previously valid class, prediction region, or learned behaviour can become invalid, irrelevant, or no longer actionable as deployment conditions change.The paper calls this task obsolescence.Source: §1 · p. 1
  • Existing responses leave obsolete outputs in place, alter exposed predictions without necessarily changing latent representations, or require data and substantial computation for retraining or fine-tuning.Unlearning and editing remove or modify behaviour but do not consider reusing obsolete structure.Source: §1 · p. 1
  • DR removes obsolete identity while reallocating retained-compatible evidence, producing repaired outputs where obsolete regions are invalid but useful obsolete structure can still support the retained task.The paper reports adaptation without full retraining while reducing obsolete behaviour and preserving retained performance across multiple task settings.Source: §1 · pp. 1–2

2 RELATED WORK

Prior approaches can retrain, unlearn, edit, or mask obsolete behaviour, but do not directly treat task obsolescence as an opportunity to repurpose useful obsolete structure. DR instead redirects that structure toward retained regions.

2 RELATED WORK

  • Retraining and fine-tuning adapt revised tasks but require training data and significant computation, while unlearning can harm retained performance.Unlearning treats unwanted behaviour primarily as information to suppress.Source: §2 · p. 2
  • Model editing and classifier repair alter targeted behaviours, but generally erase, weaken, or redirect them rather than distinguish obsolete identity from useful evidence.These methods can be cheaper than retraining.Source: §2 · p. 2
  • Representation-based methods use prototypes, distances, or feature directions, but are typically designed for recognition, generalisation, or generic editing rather than task-obsolescence repair.DR uses representation geometry to remove obsolete-supporting components, redirect useful evidence, and reconstruct outputs and latents without gradient updates.Source: §2 · p. 2
  • DR fills the identified gap by analytically redirecting useful obsolete structure toward the retained task.Source: §2 · p. 2

3 PRELIMINARIES

The paper formulates post-deployment adaptation over task-specific decision regions, some of which may become obsolete. DR seeks to remove their prediction behaviour while preserving or improving retained-region performance without updating pretrained parameters.

3 PRELIMINARIES

  • The setup uses supervised train, calibration, and test sets drawn from an underlying distribution, with a model composed of a latent representation and prediction head.The latent is the penultimate representation.Source: §3 · p. 2
  • A task-region map assigns targets to a finite set of decision regions; classification regions are output classes, while other tasks may use intervals, prototypes, anchors, or application-defined regimes.The obsolete subset is O and the remaining regions are retained as R.Source: §3 · p. 2
  • DR constructs a lightweight repaired predictor to remove behaviour associated with obsolete regions while preserving or improving performance on retained regions.The pretrained parameters remain fixed during repair and deployment adaptation.Source: §3 · p. 2

4 DEEP REPURPOSING

Deep Repurposing (DR) repairs a model analytically in latent space using fixed quantities estimated before deployment. At inference it targets obsolete predictions, redirects compatible evidence toward retained regions, and returns repaired outputs and representations.

4 DEEP REPURPOSING

  • DR has offline calibration and online inference phases, using a small calibration set to estimate region geometry, obsolete-retained directions, and residual correction prototypes.At inference, fixed quantities repair each input without gradients, optimiser steps, or fine-tuning.Source: §4 · pp. 2–3, Figure 1 · p. 3

4.1 OFFLINE CALIBRATION

  • Offline calibration fits diagonal Gaussian prototypes that estimate where retained and obsolete regions lie in latent space, and defines region evidence from model scores or calibrated latent likelihoods.The pretrained network serves as a fixed feature extractor and predictor; no gradients are computed and no parameters are updated.Source: §4.1 · p. 3
  • For each obsolete-retained pair, DR uses removal and transport directions to strip obsolete-supporting movement and redirect it within retained latent geometry.For score-head models, the removal direction follows the classifier; the transport direction comes from calibrated prototypes. The prototype-based case uses the same axis in opposite directions.Source: §4.1 · pp. 3–4
  • Residual refinement is calibrated on retained samples by comparing core-repair responsibilities with retained targets and storing reliability-adjusted destination residuals.The correction is calibrated after the core map, with a baseline for typical obsolete responsibility among retained calibration samples.Source: §4.1 · p. 4

4.2 ONLINE INFERENCE

  • Online inference activates repair for obsolete predictions, removes the obsolete-supporting component, and uses aligned retained scores and positive contrastive gains to construct retained responsibilities.When the repair mask is inactive, the original latent and retained evidence are preserved.Source: §4.2 · p. 5
  • The repaired output makes obsolete evidence non-competitive, while the returned latent is reconstructed from retained structure and kept consistent with repaired head scores when a score head is available.Without a score head, DR returns the retained posterior barycentre; pretrained parameters remain fixed throughout calibration and inference.Source: §4.2 · pp. 6–7

5 EXPERIMENTAL EVALUATION

The evaluation compares DR with unmodified, unlearning, editing, fine-tuning, and retraining approaches across diverse datasets and task types. Across these settings, DR suppresses obsolete behaviour while preserving retained utility, with results reported for classification, regression, and segmentation.

5 EXPERIMENTAL EVALUATION

  • The evaluation spans multiple approaches and datasets, with metrics covering retained performance, obsolete-output suppression, predictive reliability, representation repair, and computational cost.Datasets include classification, regression, and segmentation settings.Source: §5.1 · p. 7, §5 · p. 7

5.1 CORE RESULTS

  • Across classification datasets, DR preserves or improves retained performance while suppressing obsolete behaviour at low repair cost; Figure 3 compares obsolete probability with obsolete prediction rate.Competing methods either sacrifice retained utility or require substantially greater optimisation cost.Source: Figure 3 · p. 8, §5.1 · p. 8
  • On UTKFace, DR reduces the obsolete prediction rate from the base model’s 0.61 ± 0.01 to 0.00 ± 0.00 while retained MAE changes from 5.69 ± 0.04 to 5.68 ± 0.08.Table 1 compares retained-set error, obsolete-set prediction rate, and repair time; DR takes 16.00 ± 2.81 seconds versus 1208.13 ± 31.28 seconds for retraining.Source: Table 1 · p. 8, §5.1 · p. 8
  • On CIFAR10, DR matches retraining in 1.62 ± 0.04 seconds versus 162.17 ± 1.12 seconds; on ImageNet, both DR and the base model reach ≈81% accuracy versus ≈58% for retraining.On CIFAR100, DR preserves base-model accuracy, improves F1 from 0.79 to 0.80, and repairs in 8.25 ± 0.24 seconds.Source: §5.1 · p. 8
  • On Bike Sharing, DR reduces obsolete-region responsibility and winner rates while improving retained and obsolete-set regression errors.For light precip as the obsolete region, retained-set MAE changes from 16.34 to 16.27 and obsolete-set MAE from 14.53 to 13.72; Table 2 reports these results.Source: Table 2 · p. 9, §5.1 · p. 9
  • In the ADE20K example, DR redirects the obsolete swivel-chair subtype to the retained chair class while largely preserving surrounding segmentation structure.Figure 4 illustrates taxonomy-level repurposing from obsolete chair subtypes to the retained chair class.Source: Figure 4 · p. 9, §5.1 · p. 9
  • In deployment case studies, DR eliminates obsolete predictions on a Mars-rover classifier and a deep-sea fish classifier, with retained performance gains reported for both.The Mars cases improve retained macro-F1 by +0.097 and +0.095 in under ≈15 seconds; the fish case reports a +0.009 retained-accuracy gain in ≈7.53 seconds.Source: §5.1 · p. 9

6 CONCLUSIONS

The paper concludes that Deep Repurposing suppresses obsolete behaviour while preserving retained utility across classification, regression, and segmentation. Its stated scope depends on a well-specified obsolete region and representative calibration data, with performance depending on latent separability.

6 CONCLUSIONS

  • Across classification, regression, and segmentation, DR suppresses obsolete behaviour while preserving retained utility, redirecting obsolete structure rather than simply masking, forgetting, or retraining it away.The authors introduce DR as a post-hoc framework for adapting models when part of the task space becomes obsolete.Source: §6 · p. 9
  • DR relies on a well-specified obsolete region and representative calibration data, and its performance depends on latent separability.Source: §6 · p. 9
  • Future work will extend DR to foundation and multimodal models and develop automatic detection of obsolete behaviours.Source: §6 · p. 9

APPENDIX CONTENTS

  • The appendices cover implementation details, theoretical analysis, case studies, additional experiments and discussion, and experimental setup.Listed case studies include Mars surface classification and deep-sea fish classification.Source: APPENDIX CONTENTS · p. 13

A IMPLEMENTATION DETAILS

DR calibrates latent-region prototypes and repair directions once before deployment without updating pretrained parameters. At inference, it intervenes on inputs with obsolete evidence, repairs retained evidence, and constructs repaired predictions and latent representations.

A IMPLEMENTATION DETAILS

  • Calibration extracts pretrained-model latents and predictions, groups them by task region, and estimates a diagonal Gaussian prototype for each region without updating model parameters.This calibration stage is performed once before deployment.Source: A IMPLEMENTATION DETAILS · p. 13
  • Calibration fits obsolete-to-retained removal and transport directions, then estimates reliability-shrunk residual correction vectors and an obsolete-responsibility scale for inference.The directions identify obsolete-supporting components and specify how removed evidence may be redirected toward retained regions.Source: A IMPLEMENTATION DETAILS · p. 13
  • At inference, DR intervenes only when the strongest region evidence is obsolete, limiting modification of inputs already assigned to retained regions.It selects the strongest obsolete region and retained alternative, strips the obsolete-supporting component, and combines original retained, prototype, and contrastive repair evidence.Source: A IMPLEMENTATION DETAILS · pp. 13–14
  • DR normalizes repaired evidence over retained regions and applies residual refinement gated by the input’s relative obsolete responsibility.For inputs not triggering repair, it retains the original latent and retained evidence.Source: A IMPLEMENTATION DETAILS · p. 14
  • The output construction depends on task type, and DR also constructs a repaired latent using retained-head consistency or retained-prototype geometry.An SVD-based Moore–Penrose pseudoinverse provides a minimum-norm score-consistent solution when well posed or a least-squares projection when rank-deficient or overdetermined.Source: A IMPLEMENTATION DETAILS · pp. 14–15

B THEORETICAL ANALYSIS

The theoretical analysis shows how DR suppresses obsolete evidence while transporting representations toward retained structure without undoing the retained-side margin under stated compatibility conditions. The case-study discussion motivates deployment adaptation and describes Mars rover settings where hardware or taxonomy changes make classes obsolete.

B THEORETICAL ANALYSIS

  • DR’s classifier-induced direction suppresses obsolete-relative-to-retained evidence, while prototype transport restores retained structure without requiring the two directions to be collinear.The retained-side margin is preserved when transport does not increase the obsolete-retained margin, or when transport strength is capped to limit that increase.Source: B THEORETICAL ANALYSIS · pp. 15, 17
  • The repaired all-region softmax assigns every obsolete region lower probability than every retained region, so its argmax selects a retained region.The score construction uses the weakest retained region as a reference and a stabiliser to keep scores finite.Source: B THEORETICAL ANALYSIS · p. 17

C CASE STUDIES

  • The case studies assess DR in operational settings where class relevance can shift organically, data is limited, and retraining from scratch may not be an option.Source: C CASE STUDIES · p. 18

CASE STUDY 1: MARS SURFACE CLASSIFICATION

  • The Mars rover study uses an imbalanced MSL image classifier and tests retiring DRT-related classes after hardware or mission changes, as well as removing horizon when it harms retained macro-F1.The two variants represent hardware-induced obsolescence and taxonomy-induced obsolescence.Source: CASE STUDY 1: MARS SURFACE… · p. 18, Figure 6 · p. 19

HARDWARE-INDUCED OBSOLESCENCE

Across the Mars rover cases, DR removes obsolete outputs while improving retained macro-F1; adaptation takes seconds rather than minutes. The FishNet case uses deployment-time uncertainty to identify an unreliable class, which DR suppresses while preserving the remaining taxonomy.

HARDWARE-INDUCED OBSOLESCENCE

  • DR reduced obsolete DRT prediction rate from 0.672 to 0.000 and improved retained Macro-F1 from 0.759 to 0.856.The raw probability assigned to obsolete DRT classes fell from 0.398 to approximately zero; NLL decreased from 0.887 to 0.579, while ECE increased from 0.157 to 0.186. DRT samples were redirected toward nearby retained classes rather than simply discarded.Source: HARDWARE-INDUCED OBSOLESCENCE · pp. 19–20, Table 3 · p. 20, Table 4 · p. 20

TAXONOMY-INDUCED OBSOLESCENCE

COMPUTATIONAL EFFICIENCY

  • DR adapted the rover classifier in 6.6 seconds on an A100, compared with 310.7 seconds for fine-tuning, a reported 47× speedup.DR used approximately 4.1GB VRAM, while fine-tuning required up to 14.5GB and exhausted the available memory budget in the budget-hardware setting.Source: COMPUTATIONAL EFFICIENCY · p. 22

CASE STUDY 2: DEEP-SEA FISH CLASSIFICATION

  • On FishNet, uncertainty estimation identified poorly performing, high-uncertainty Perciformes for removal; DR reduced its obsolete prediction rate from 0.394 to 0.000 and improved retained accuracy from 0.687 to 0.696.The obsolete class also leaked into retained samples before repair: its retained obsolete prediction rate fell from 0.026 to 0.000. DR completed adaptation in 7.53 ± 0.20 seconds on an A40 GPU.Source: CASE STUDY 2: DEEP-SEA FISH… · pp. 23–25, Table 6 · p. 24

D ADDITIONAL EXPERIMENTS

The additional experiments extend evaluation beyond headline results to diagnostics, robustness checks, and component ablations.

D ADDITIONAL EXPERIMENTS

  • The additional experiments report full results, calibration and confidence diagnostics, representation behaviour, retained performance across difficulty levels, and extended regression and segmentation results.They also compare DR with logit suppression and ablate its main components, calibration-set size, and residual refinement.Source: D ADDITIONAL EXPERIMENTS · p. 25

FULL RESULTS

Across classification, regression, and segmentation tasks, Deep Repurposing (DR) suppresses obsolete outputs while maintaining retained utility. Diagnostics also show reduced obsolete probability and latent structure, with a stronger repair than logit suppression alone.

FULL RESULTS

  • Across classification datasets, DR preserves retained utility at or above the base model while reducing obsolete prediction rate and probability to 0.00 on Do.On CIFAR10, it matches retraining in retained accuracy and F1, with average repair times of 1.62 seconds for DR and 162.17 seconds for retraining.Source: FULL RESULTS · pp. 25–26
  • On retained classification data, DR matches CIFAR10 ECE at 0.03 and improves NLL by −0.03 on average, while calibration instability occurs for some suppression methods.Obsolete-split NLL needs different interpretation because evaluation uses original labels, where obsolete classes remain targets; read it alongside obsolete prediction rates.Source: FULL RESULTS · p. 26, Table 8 · p. 26
  • On retained classification data, DR keeps confidence/entropy close to the base model, while obsolete-sample confidence falls and entropy rises on CIFAR10 and CIFAR100.For CIFAR100 Do, confidence/entropy shifts from 0.78/1.09 to 0.29/3.25; DR’s retained CIFAR10 values are 0.94/0.18 versus 0.93/0.19 for the base model.Source: FULL RESULTS · pp. 26–27, Table 9 · p. 27
  • On ImageNet, DR records 0.11 near-obsolete-centroid rate and 0.23 obsolete-direction score, while preserving base retained accuracy.On CIFAR100, DR’s retained accuracy is 0.80, compared with 0.62 for Boundary Shrink despite a similar centroid rate of 0.05.Source: FULL RESULTS · pp. 27–28, Table 10 · p. 28
  • DR maintains retained-set utility on difficult examples: CIFAR10 Q4 accuracy/F1 rises from 0.61/0.56 to 0.64/0.64, and CIFAR100 Q4 rises from 0.19/0.19 to 0.20/0.20.Comparative methods sometimes degrade harder retained examples; fine-tuning and retraining improve CIFAR10 Q4 F1 to 0.67 and 0.66, respectively, with optimization cost.Source: FULL RESULTS · pp. 28–29, Table 11 · p. 29
  • On UTKFace regression, DR changes retained Dr MAE from 5.69 to 5.68, keeps Within 10 at 0.830, and reduces obsolete prediction rate to 0.000 on both splits.Retraining lowers Dr MAE to 4.46 and raises Within 10 to 0.895, but averages 1208.13 seconds.Source: FULL RESULTS · p. 30, Table 12 · p. 30
  • On ADE20K segmentation, DR reduces obsolete prediction rate and effective obsolete probability to zero while retained mIoU stays approximately 0.36.On the obsolete-containing split, retained-pixel accuracy rises from 0.74 to 0.75; all-pixel accuracy is evaluated against original labels that still include obsolete chair subtypes.Source: FULL RESULTS · p. 30, Table 13 · p. 30

COMPARISON TO LOGIT SUPPRESSION

  • On CIFAR10, logit suppression and DR both remove obsolete predictions after post-processing, but raw obsolete probability is 0.88 ± 0.05 under suppression and 0.00 ± 0.00 under DR.DR also reduces the near-obsolete-centroid rate from 0.90 ± 0.04 to 0.04 ± 0.05 and takes 1.16 ± 0.04 seconds; logit suppression is essentially cost-free and suffices if only hiding the final label is required.Source: COMPARISON TO LOGIT SUPPRESSION · pp. 30–31, Table 14 · p. 31

ABLATION STUDY

The ablations find that DR’s retained performance is generally stable across hyperparameter settings, while stronger repair can trade off against retained performance or compute. The discussion distinguishes DR’s operational post-hoc repair from federated learning, privacy-oriented unlearning, and continual learning.

ABLATION STUDY

  • Across stripping-margin values, retained F1 stays nearly flat at roughly 0.9055–0.9057, although obsolete-split raw confidence rises mildly as m increases.The overlapping uncertainty bands indicate low sensitivity to m.Source: ABLATION STUDY · p. 31, Figure 13 · p. 32
  • As β increases, retained F1 stays effectively unchanged while retained ECE decreases, and obsolete-split raw confidence falls as entropy rises.The passage attributes the calibration and suppression pattern to retained prototype evidence.Source: ABLATION STUDY · p. 32, Figure 14 · p. 32
  • Increasing λ reduces near-obsolete-centroid rate and obsolete-direction score, but large values slightly lower retained F1 and sharply raise obsolete-split raw confidence.Small-to-intermediate λ values preserve retained performance while improving obsolete-geometry suppression.Source: ABLATION STUDY · p. 32, Figure 15 · p. 33
  • Across calibration-set sizes, retained F1 is effectively flat and near-obsolete-centroid rate stays low, while larger sets reduce obsolete-split raw confidence and increase repair time.CIFAR10 repair takes approximately 0.25 seconds at 500 or fewer samples and approximately 3.0 seconds at 10,000 samples.Source: ABLATION STUDY · pp. 32–33, Figure 16 · p. 33
  • On the MSL case study, both DR variants reduce obsolete prediction rate and obsolete probability mass from base-model averages of 0.672 and 0.398 to 0.000.Residual refinement mainly improves retained utility: accuracy rises to 0.884 and macro-F1 to 0.856, compared with 0.803 accuracy and 0.792 macro-F1 without refinement.Source: ABLATION STUDY · p. 33, Table 15 · p. 33

FEDERATED LEARNING

  • DR complements federated learning: it locally repairs an already deployed classifier without waiting for a federated training cycle or transmitting a replacement model.Federated retraining may remain useful for longer-term global maintenance but requires communication, coordination, validation, and redeployment.Source: FEDERATED LEARNING · p. 34

DEEP UNLEARNING AND PRIVACY

  • DR is operational repair, not certified deletion: pretrained parameters remain fixed, and obsolete regions may still have shaped learned representations.For personal data, consent withdrawal, or regulatory removal, the passage identifies unlearning, retraining, or another deletion mechanism as appropriate.Source: DEEP UNLEARNING AND PRIVACY · pp. 34–35

CONTINUAL LEARNING

  • Unlike continual learning, DR does not learn task sequences or new classes; it repairs a fixed model when part of its existing output space becomes obsolete.It keeps pretrained parameters fixed and uses a calibration-derived analytic repair map, whereas continual-learning approaches typically require updates or optimization.Source: CONTINUAL LEARNING · p. 35

F EXPERIMENTAL SETUP

The evaluation fixes obsolete classes or regions before comparing methods, using exhaustive selection for CIFAR-10 and stratified difficulty coverage for larger label spaces. Regression experiments test both target-interval and semantic-region obsolescence.

SELECTION OF OBSOLETE CLASSES/REGIONS

  • CIFAR-10 uses an exhaustive sweep, treating each of its ten classes as obsolete once.The small label space avoids a sampling decision.Source: SELECTION OF OBSOLETE CLASSES/REGIONS · p. 35
  • For large-label datasets, a reliability-risk ranking selects six obsolete classes spanning easy, medium, and hard cases instead of an exhaustive sweep.The score combines standardized class-level NLL, ECE, entropy, and accuracy; the same selected classes are used across methods.Source: SELECTION OF OBSOLETE CLASSES/REGIONS · pp. 35–36
  • For UTKFace, the oldest age interval, 60–120, is fixed as obsolete because it scores worst on several calibration metrics and is also highest-ranked on test.All remaining age intervals are retained.Source: SELECTION OF OBSOLETE CLASSES/REGIONS · p. 37
  • For Bike Sharing, light precipitation is obsolete as the hardest populated weather regime, while clear, mist, and heavy precipitation remain retained.The selection is supported by the largest calibration MAE and 90th-percentile absolute error, with the same pattern on validation.Source: SELECTION OF OBSOLETE CLASSES/REGIONS · p. 37
  • Both regression regions are selected using the base model and a fixed calibration protocol before methods are applied, with shared test splits across methods.This tests target-associated obsolescence in UTKFace and semantic-region obsolescence in Bike Sharing.Source: SELECTION OF OBSOLETE CLASSES/REGIONS · p. 37

DATASETS

The experimental suite spans diverse benchmarks, model families, and repair baselines, with fixed training checkpoints shared across comparisons. DR’s runtime remains seconds-scale in the reported experiments.

DATASETS

  • The benchmark suite covers established computer-vision and tabular tasks across varied domains and complexity levels.The section provides dataset details and representative examples.Source: DATASETS · p. 37

MODELS

  • Experiments use image classifiers, semantic segmentation, and tabular regression models, with dataset-specific checkpoints fixed across repair methods.Models are fine-tuned on the complete original task before classes are designated obsolete.Source: MODELS · p. 39

COMPARATIVE APPROACHES

  • Comparators include fine-tuning and retraining alongside unlearning and editing methods; DR performs parameter-free latent-space repair on a fixed pretrained model.DR estimates region prototypes and repair directions during calibration, then analytically removes and redistributes obsolete evidence at inference.Source: COMPARATIVE APPROACHES · pp. 40–41

TRAINING DETAILS

  • Experiments use a consistent training configuration and evaluate ten random seeds per obsolete class.Runs used a performance GPU cluster with a maximum of three Nvidia A40 GPUs.Source: TRAINING DETAILS · p. 41

RUNTIME AND COMPLEXITY

  • ≈1.62s on CIFAR10, with DR remaining seconds-scale across substantially larger models and task spaces.Inference operations scale with prediction locations and region counts; contrastive transport is evaluated only where the repair mask activates.Source: RUNTIME AND COMPLEXITY · pp. 41–42
Loading 2610.01453v1…