Repurposing Obsolete Representations for Post-Deployment Adaptation

Daniel Bethell; Charmaine Barker; Simos Gerasimou

cs.LGOct 1, 2026

Repurposing Obsolete Representations for Post-Deployment Adaptation

Authors: Daniel Bethell, Charmaine Barker, Simos Gerasimou

Organizations: Department of Computer Science, University of York, UK · Department of Elect. Eng., and Computer Science and Eng., Cyprus University of Technology, Cyprus

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 $60\times$ faster than competing unlearning methods.

Figures & tables

Figure 1: Overview of Deep Repurposing (DR). Offline calibration estimates retained and obsolete latent geometry and residual correction quantities using a frozen pretrained model. At inference, DR detects obsolete predictions, removes obsolete-supporting components, repurposes evidence towards the best-suited retained regions, applies residual refinement, and returns repaired predictions and latent representations.

Figure 2: Classification repair trade-off across all datasets. Each panel compares retained-set F1 on $D_{\mathrm{r}}$ and obsolete-set F1 on $D_{\mathrm{o}}$ against repair time.

Figure 3: Obsolete-output suppression across classification datasets. Each panel compares the raw probability assigned to obsolete classes on $D_{\mathrm{o}}$ against the obsolete prediction rate on $D_{\mathrm{o}}$ .

Method	$D_{\mathrm{r}}$ MAE $\downarrow$	$D_{\mathrm{o}}$ Obs. Rate $\downarrow$	Time (s) $\downarrow$
Base Model	$5.69\pm 0.04$	$0.61\pm 0.01$	–
Finetune	$5.28\pm 0.07$	$0.38\pm 0.03$	$282.05\pm 100.70$
Retrain	$4.46\pm 0.03$	$0.00\pm 0.00$	$1208.13\pm 31.28$
DR	$5.68\pm 0.08$	$0.00\pm 0.00$	$16.00\pm 2.81$

Table 1: Regression trade-off results on the UTKFace dataset where the obsolete region is tied to the target, comparing retained set error, obsolete-set prediction rate and repair time.

		Predictive performance			Semantic-region performance
Method	Split	MAE $\downarrow$	RMSE $\downarrow$	$R^{2}$ $\uparrow$	$q_{\mathrm{obs}}$ $\downarrow$	Obs. win $\downarrow$
Base	$D_{\mathrm{r}}$	$16.34\pm 0.36$	$32.12\pm 1.12$	$0.956\pm 0.002$	$0.430\pm 0.120$	$0.447\pm 0.131$
Base	$D_{\mathrm{o}}$	$14.53\pm 1.50$	$35.92\pm 3.99$	$0.897\pm 0.019$	$0.562\pm 0.154$	$0.589\pm 0.177$
DR	$D_{\mathrm{r}}$	$16.27\pm 0.48$	$31.35\pm 1.62$	$0.958\pm 0.003$	$0.283\pm 0.206$	$0.237\pm 0.287$
DR	$D_{\mathrm{o}}$	$13.72\pm 1.74$	$31.94\pm 4.68$	$0.921\pm 0.019$	$0.376\pm 0.255$	$0.304\pm 0.375$

Table 2: Semantic-region repurposing results on the Bike Sharing dataset. The obsolete region is light_precip ; retained regions are clear , mist , and heavy_precip .

q_{\mathrm{obs}}

denotes obsolete-region responsibility, and Obs. win denotes the fraction of samples assigned to the obsolete region.

Figure 4: ADE20K segmentation example showing taxonomy-level repurposing from obsolete chair subtypes { armchair , seat , swivel chair , stool } to the retained chair class.

Appendix figures & tables30 assets

Supplementary material from the paper’s appendix.

Appendix

\mathcal{M}(x)=\mathbf{1}\left[\arg\max_{c\in\mathcal{O}\cup\mathcal{R}}e_{c}(z)\in\mathcal{O}\right]

Appendix

Algorithm 2 DeepRepurposing - Inference

Figure 5: Example images from the MSL rover classification dataset that contains visually diverse mission-relevant categories.

Figure 6: Overview of the Mars rover Deep Repurposing case study. A deployed MSL image classifier is adapted under two mission-phase changes: hardware-induced obsolescence, where a DRT-related class becomes irrelevant after tool damage, retirement, or mission-phase completion; and taxonomy-induced obsolescence, where the horizon class is found to harm retained macro-F1 performance. Deep Repurposing suppresses obsolete output behaviour and reallocates learned structure toward the retained mission classes.

Metric	Base Model	DR	$\Delta$
$D_{r}$ accuracy	0.803 $\pm$ 0.001	0.884 $\pm$ 0.001	+0.081
$D_{r}$ macro-F1	0.759 $\pm$ 0.002	0.856 $\pm$ 0.002	+0.097
$D_{r}$ NLL	0.887 $\pm$ 0.002	0.579 $\pm$ 0.002	-0.308
$D_{r}$ ECE	0.157 $\pm$ 0.001	0.186 $\pm$ 0.001	+0.029
$D_{r}$ obs prediction rate	0.001 $\pm$ 0.000	0.000 $\pm$ 0.000	-0.001
$D_{r}$ raw obs probability	0.021 $\pm$ 0.000	0.000 $\pm$ 0.000	-0.021

Appendix

Table 3: Results across 10 runs (10 random seeds) for the hardware-induced obsolescence experiment, where DRT-related classes are treated as obsolete.

Figure 7: Qualitative visualisation of DRT repurposing from an example random seed. Left: The original representation of all MSL classes. Middle: The repurposing movement from obsolete DRT outputs toward retained alternatives (represented by arrows). Right: The new/repaired representation of all MSL classes (obsolete class still visible for qualitative purposes).

Nearest retained class	Before DR	After DR
Portion box	71 / 143 (49.7%)	54 / 143 (37.8%)
Ground	42 / 143 (29.4%)	20 / 143 (14.0%)
Wheel	27 / 143 (18.9%)	34 / 143 (23.8%)
Scoop	1 / 143 (0.7%)	27 / 143 (18.9%)
Drill	0 / 143 (0.0%)	6 / 143 (4.2%)
MAHLI calibration target	2 / 143 (1.4%)	2 / 143 (1.4%)

Appendix

Table 4: Nearest retained-class distribution for obsolete DRT samples before and after Deep Repurposing from an example random seed.

Metric	Base Model	DR	$\Delta$
$D_{r}$ accuracy	0.784 $\pm$ 0.001	0.884 $\pm$ 0.001	+0.100
$D_{r}$ macro-F1	0.739 $\pm$ 0.002	0.834 $\pm$ 0.002	+0.095
$D_{r}$ NLL	0.933 $\pm$ 0.002	0.658 $\pm$ 0.002	-0.274
$D_{r}$ ECE	0.151 $\pm$ 0.000	0.211 $\pm$ 0.001	+0.060
$D_{r}$ obs prediction rate	0.063 $\pm$ 0.001	0.000 $\pm$ 0.000	-0.063
$D_{r}$ raw obs probability	0.029 $\pm$ 0.000	0.000 $\pm$ 0.000	-0.029

Appendix

Table 5: Results across 10 runs (10 random seeds) for the taxonomy-induced obsolescence experiment, where the horizon class is obsolete.

Figure 8: Qualitative visualisation of horizon repurposing from an example random seed. Left: The original representation of all MSL classes. Middle: The repurposing movement from obsolete horizon outputs toward retained alternatives (represented by arrows). Right: The new/repaired representation of all MSL classes (obsolete class still visible for qualitative purposes).

Figure 9: Example images from the FishNet classification dataset that contains visually diverse orders.

Figure 10: Overview of the uncertainty-triggered FishNet Deep Repurposing case study. During deployment, an autonomous underwater vehicle repeatedly observes new fish imagery, predicts fish orders, and estimates predictive uncertainty. Persistently high uncertainty and poor performance for the ’Perciformes’ class triggers Deep Repurposing, which suppresses the obsolete class output and reallocates learned structure toward retained fish orders while preserving retained performance and reducing obsolete predictions.

Figure 11: Uncertainty estimates and accuracy for FishNet orders using a last-layer Laplace approximation; larger bubbles indicate orders with more samples.

Metric	Base Model	DR	$\Delta$
$D_{r}$ accuracy	0.687 $\pm$ 0.000	0.696 $\pm$ 0.000	+0.009
$D_{r}$ macro-F1	0.579 $\pm$ 0.000	0.586 $\pm$ 0.002	+0.007
$D_{r}$ NLL	1.169 $\pm$ 0.0020	1.121 $\pm$ 0.000	-0.048
$D_{r}$ ECE	0.074 $\pm$ 0.000	0.082 $\pm$ 0.000	+0.008
$D_{r}$ obs prediction rate	0.026 $\pm$ 0.000	0.000 $\pm$ 0.000	-0.026
$D_{r}$ raw obs probability	0.039 $\pm$ 0.000	0.000 $\pm$ 0.000	-0.039

Appendix

Table 6: Results across 10 runs (10 random seeds) for the uncertainty-induced obsolescence experiment, where the class with the highest uncertainty and lowest performance (Perciformes) is obsolete.

Figure 12: Per-class effect of Deep Repurposing on FishNet after removing the uncertainty-triggered obsolete class (Perciformes). Each point shows the change in accuracy and Laplace NLL for a retained fish order after repurposing, with the dashed lines indicating no change. The inset shows the full-scale view including Perciformes, whose accuracy drops sharply and whose NLL increases substantially after removal, reflecting the intended suppression of the obsolete class.

	$\mathbf{D_{\mathrm{r}}}$				$\mathbf{D_{\mathrm{o}}}$				Time
Method	Acc. $\uparrow$	F1 $\uparrow$	Obs. Rate $\downarrow$	Obs. Prob. $\downarrow$	Acc. $\downarrow$	F1 $\downarrow$	Obs. Rate $\downarrow$	Obs. Prob. $\downarrow$	Unlearn (s) $\downarrow$
CIFAR10
Base Model	0.90 $\pm$ 0.00	0.82 $\pm$ 0.01	0.01 $\pm$ 0.00	0.01 $\pm$ 0.00	0.90 $\pm$ 0.04	0.12 $\pm$ 0.03	0.90 $\pm$ 0.04	0.88 $\pm$ 0.05	–
SSD	0.88 $\pm$ 0.01	0.82 $\pm$ 0.01	0.00 $\pm$ 0.00	0.01 $\pm$ 0.00	0.36 $\pm$ 0.15	0.05 $\pm$ 0.02	0.36 $\pm$ 0.15	0.33 $\pm$ 0.13	2.32 $\pm$ 0.49
LFSSD	0.88 $\pm$ 0.01	0.83 $\pm$ 0.02	0.00 $\pm$ 0.00	0.01 $\pm$ 0.00	0.40 $\pm$ 0.20	0.05 $\pm$ 0.02	0.40 $\pm$ 0.20	0.36 $\pm$ 0.19	4.30 $\pm$ 1.34
Boundary Shrink	0.79 $\pm$ 0.03	0.76 $\pm$ 0.03	0.00 $\pm$ 0.00	0.00 $\pm$ 0.00	0.07 $\pm$ 0.01	0.01 $\pm$ 0.00	0.07 $\pm$ 0.01	0.07 $\pm$ 0.01	13.12 $\pm$ 0.31

Appendix

Table 7: Full classification results across all datasets reporting retained performance, obsolete-output suppression, and repair time.

	CIFAR10				CIFAR100
Method	$\mathbf{D_{r}}$ NLL $\downarrow$	$\mathbf{D_{r}}$ ECE $\downarrow$	$\mathbf{D_{o}}$ NLL $\downarrow$	$\mathbf{D_{o}}$ ECE $\downarrow$	$\mathbf{D_{r}}$ NLL $\downarrow$	$\mathbf{D_{r}}$ ECE $\downarrow$	$\mathbf{D_{o}}$ NLL $\downarrow$	$\mathbf{D_{o}}$ ECE $\downarrow$
Base Model	0.30 $\pm$ 0.01	0.03 $\pm$ 0.00	0.30 $\pm$ 0.13	0.04 $\pm$ 0.02	0.80 $\pm$ 0.01	0.03 $\pm$ 0.01	0.83 $\pm$ 0.64	0.14 $\pm$ 0.06
SSD	0.36 $\pm$ 0.04	0.02 $\pm$ 0.00	2.07 $\pm$ 0.62	0.32 $\pm$ 0.10	1.51 $\pm$ 0.82	0.21 $\pm$ 0.07	5.73 $\pm$ 0.62	0.16 $\pm$ 0.05
LFSSD	0.36 $\pm$ 0.04	0.03 $\pm$ 0.00	2.25 $\pm$ 0.92	0.32 $\pm$ 0.14	0.86 $\pm$ 0.04	0.07 $\pm$ 0.03	4.61 $\pm$ 0.39	0.35 $\pm$ 0.10
Boundary Shrink	0.72 $\pm$ 0.13	0.08 $\pm$ 0.03	4.82 $\pm$ 0.23	0.60 $\pm$ 0.02	2.12 $\pm$ 0.25	0.35 $\pm$ 0.05	4.78 $\pm$ 0.07	0.10 $\pm$ 0.04
Boundary Expand	0.67 $\pm$ 0.06	0.02 $\pm$ 0.00	3.59 $\pm$ 0.16	0.43 $\pm$ 0.01	1.81 $\pm$ 0.29	0.19 $\pm$ 0.03	4.98 $\pm$ 0.05	0.16 $\pm$ 0.03

Appendix

Table 8: Classification calibration results across all datasets reporting negative log-likelihood (NLL) and expected calibration error (ECE) on retained and obsolete test splits.

	CIFAR10				CIFAR100
Method	$\mathbf{D_{r}}$ Conf. $\downarrow$	$\mathbf{D_{r}}$ Ent. $\downarrow$	$\mathbf{D_{o}}$ Conf. $\downarrow$	$\mathbf{D_{o}}$ Ent. $\downarrow$	$\mathbf{D_{r}}$ Conf. $\downarrow$	$\mathbf{D_{r}}$ Ent. $\downarrow$	$\mathbf{D_{o}}$ Conf. $\downarrow$	$\mathbf{D_{o}}$ Ent. $\downarrow$
Base Model	0.93 $\pm$ 0.00	0.19 $\pm$ 0.01	0.93 $\pm$ 0.02	0.19 $\pm$ 0.07	0.79 $\pm$ 0.01	1.11 $\pm$ 0.00	0.78 $\pm$ 0.09	1.09 $\pm$ 0.22
SSD	0.90 $\pm$ 0.02	0.31 $\pm$ 0.05	0.64 $\pm$ 0.07	1.05 $\pm$ 0.22	0.48 $\pm$ 0.21	2.60 $\pm$ 0.85	0.16 $\pm$ 0.04	3.92 $\pm$ 0.19
LFSSD	0.91 $\pm$ 0.01	0.26 $\pm$ 0.03	0.70 $\pm$ 0.08	0.86 $\pm$ 0.22	0.73 $\pm$ 0.03	1.42 $\pm$ 0.17	0.36 $\pm$ 0.10	2.90 $\pm$ 0.55
Boundary Shrink	0.86 $\pm$ 0.01	0.40 $\pm$ 0.03	0.67 $\pm$ 0.01	0.97 $\pm$ 0.04	0.27 $\pm$ 0.03	3.65 $\pm$ 0.11	0.14 $\pm$ 0.02	4.05 $\pm$ 0.06
Boundary Expand	0.80 $\pm$ 0.02	0.63 $\pm$ 0.05	0.55 $\pm$ 0.01	1.31 $\pm$ 0.02	0.43 $\pm$ 0.04	2.87 $\pm$ 0.16	0.20 $\pm$ 0.01	3.70 $\pm$ 0.06

Appendix

Table 9: Confidence and predictive entropy results across all classification datasets on retained and obsolete test splits.

CIFAR10
	AUC		$\mathbf{D_{\mathrm{r}}}$		$\mathbf{D_{\mathrm{o}}}$
Method	MIA $\downarrow$	Probe $\downarrow$	Near Obs. Centroid Rate $\downarrow$	Obs. Direction $\downarrow$	Near Obs. Centroid Rate $\downarrow$	Obs. Direction $\downarrow$
Base Model	–	–	–	–	–	–
SSD	0.52 $\pm$ 0.02	0.98 $\pm$ 0.01	0.00 $\pm$ 0.00	-0.01 $\pm$ 0.01	0.40 $\pm$ 0.15	0.26 $\pm$ 0.05
LFSSD	0.51 $\pm$ 0.02	0.98 $\pm$ 0.01	0.00 $\pm$ 0.01	-0.03 $\pm$ 0.02	0.44 $\pm$ 0.22	0.28 $\pm$ 0.11
Boundary Shrink	0.52 $\pm$ 0.02	0.96 $\pm$ 0.01	0.00 $\pm$ 0.00	-0.07 $\pm$ 0.01	0.10 $\pm$ 0.03	0.08 $\pm$ 0.02

Appendix

Table 10: Representation-level classification results across all datasets on retained and obsolete test splits. Metrics report membership-inference attack AUROC, linear-probe AUROC, proximity to obsolete centroids, and alignment with obsolete-retained directions.

	$\mathbf{D_{\mathrm{r}}}$ Accuracy				$\mathbf{D_{\mathrm{r}}}$ F1
Method	Q1 $\uparrow$	Q2 $\uparrow$	Q3 $\uparrow$	Q4 $\uparrow$	Q1 $\uparrow$	Q2 $\uparrow$	Q3 $\uparrow$	Q4 $\uparrow$
CIFAR10
Base Model	1.00 $\pm$ 0.00	1.00 $\pm$ 0.00	1.00 $\pm$ 0.00	0.61 $\pm$ 0.02	1.00 $\pm$ 0.00	1.00 $\pm$ 0.00	1.00 $\pm$ 0.00	0.56 $\pm$ 0.02
SSD	0.99 $\pm$ 0.00	0.99 $\pm$ 0.01	0.96 $\pm$ 0.02	0.58 $\pm$ 0.03	0.99 $\pm$ 0.00	0.98 $\pm$ 0.01	0.95 $\pm$ 0.02	0.53 $\pm$ 0.03
LFSSD	0.99 $\pm$ 0.00	0.99 $\pm$ 0.01	0.96 $\pm$ 0.02	0.57 $\pm$ 0.03	0.99 $\pm$ 0.00	0.99 $\pm$ 0.01	0.95 $\pm$ 0.01	0.26 $\pm$ 0.22
Boundary Shrink	0.98 $\pm$ 0.01	0.93 $\pm$ 0.03	0.81 $\pm$ 0.06	0.44 $\pm$ 0.05	0.97 $\pm$ 0.01	0.92 $\pm$ 0.03	0.80 $\pm$ 0.05	0.43 $\pm$ 0.04

Appendix

Table 11: Retained-set classification results by difficulty quartile across all datasets. Results report accuracy and F1 from the easiest retained examples (Q1) to the hardest retained examples (Q4).

Method	Split	MAE $\downarrow$	RMSE $\downarrow$	Within 5 $\uparrow$	Within 10 $\uparrow$	Obs. rate $\downarrow$	Repair (s) $\downarrow$
Base model	$D_{\mathrm{r}}$	$5.69\pm 0.06$	$7.70\pm 0.10$	$0.558\pm 0.006$	$0.830\pm 0.003$	$0.014\pm 0.002$	–
Base model	$D_{\mathrm{o}}$	$10.11\pm 0.61$	$13.34\pm 0.79$	$0.339\pm 0.020$	$0.607\pm 0.029$	$0.607\pm 0.011$	–
Fine-tuning	$D_{\mathrm{r}}$	$5.28\pm 0.10$	$7.16\pm 0.13$	$0.587\pm 0.011$	$0.855\pm 0.008$	$0.003\pm 0.001$	$282.05\pm 140.76$
Fine-tuning	$D_{\mathrm{o}}$	$14.86\pm 0.79$	$17.21\pm 0.74$	$0.111\pm 0.024$	$0.317\pm 0.047$	$0.379\pm 0.040$	$282.05\pm 140.76$
Retraining	$D_{\mathrm{r}}$	$4.46\pm 0.04$	$6.26\pm 0.04$	$0.668\pm 0.004$	$0.895\pm 0.004$	$0.000\pm 0.000$	$1208.13\pm 43.73$
Retraining	$D_{\mathrm{o}}$	$19.68\pm 1.07$	$22.21\pm 0.98$	$0.032\pm 0.021$	$0.176\pm 0.044$	$0.000\pm 0.000$	$1208.13\pm 43.73$

Appendix

Table 12: UTKFace label-region regression results reporting predictive error, within-threshold accuracy, obsolete-region suppression, and repair time.

Method	Split	Pixel Acc. $\uparrow$	Ret. Pixel Acc. $\uparrow$	Ret. mIoU $\uparrow$	NLL $\downarrow$	ECE $\downarrow$	Obs. Pred. $\downarrow$	Eff. Obs. Prob. $\downarrow$
Base	$D_{\mathrm{r}}$	$0.77\pm 0.00$	$0.77\pm 0.00$	$0.36\pm 0.00$	$\mathbf{0.83\pm 0.00}$	$0.06\pm 0.00$	$0.00\pm 0.00$	$0.00\pm 0.00$
DR	$D_{\mathrm{r}}$	$\mathbf{0.78\pm 0.01}$	$\mathbf{0.78\pm 0.01}$	$\mathbf{0.36\pm 0.01}$	$0.84\pm 0.03$	$\mathbf{0.06\pm 0.00}$	$\mathbf{0.00\pm 0.00}$	$\mathbf{0.00\pm 0.00}$
Base	$D_{\mathrm{o}}$	$\mathbf{0.72\pm 0.00}$	$0.74\pm 0.00$	$\mathbf{0.25\pm 0.00}$	$0.95\pm 0.00$	$\mathbf{0.05\pm 0.00}$	$0.07\pm 0.00$	$0.06\pm 0.00$
DR	$D_{\mathrm{o}}$	$0.68\pm 0.02$	$\mathbf{0.75\pm 0.01}$	$0.25\pm 0.00$	–	$0.06\pm 0.00$	$\mathbf{0.00\pm 0.00}$	$\mathbf{0.00\pm 0.00}$

Appendix

Table 13: ADE20K semantic segmentation results. The obsolete classes are chair subtypes

\{\texttt{armchair},\texttt{seat},\texttt{swivel chair},\texttt{stool}\}

, while all other ADE20K classes are retained.

	$\mathbf{D_{\mathrm{r}}}$ utility		$\mathbf{D_{\mathrm{o}}}$ obsolete-output behaviour
Method	Acc. $\uparrow$	F1 $\uparrow$	Obs. Rate $\downarrow$	Raw Obs. Prob. $\downarrow$	Unlearn (s) $\downarrow$
Logit Suppression	$0.91\pm 0.01$	$0.91\pm 0.01$	$0.00\pm 0.00$	$0.88\pm 0.05$	$0.00\pm 0.00$
Deep Repurposing	$0.91\pm 0.01$	$0.91\pm 0.01$	$0.00\pm 0.00$	$\mathbf{0.00\pm 0.00}$	$1.16\pm 0.04$

Appendix

Table 14: Diagnostic comparison between Logit Suppression (LS) and Deep Repurposing (DR) on CIFAR10. Although both methods remove obsolete predictions after post-processing, LS leaves substantial obsolete probability mass and obsolete latent structure in the underlying model. Lower is better for obsolete probability, centroid rate, direction score, confidence, and unlearning time.

Figure 13: Ablation of the stripping margin $m$ across various values.

Figure 14: Ablation of the Bayes weight $\beta$ across various values.

Figure 15: Ablation of the contrastive repair weight $\lambda$ across various values.

Figure 16: Ablation of the size of the calibration dataset $|\mathcal{D}_{cal}|$ across various values.

Method	$D_{\mathrm{r}}$ Acc. $\uparrow$	$D_{\mathrm{r}}$ F1 $\uparrow$	$D_{\mathrm{r}}$ NLL $\downarrow$	$D_{\mathrm{r}}$ ECE $\downarrow$	$D_{\mathrm{o}}$ Conf. $\downarrow$	$D_{\mathrm{o}}$ Ent. $\uparrow$	Obs. Rate $\downarrow$	Obs. Prob. $\downarrow$
Base	$0.803\pm 0.001$	$0.759\pm 0.002$	$0.887\pm 0.002$	$0.157\pm 0.001$	$0.433\pm 0.003$	$2.092\pm 0.008$	$0.672\pm 0.007$	$0.398\pm 0.003$
DR (w/o ref)	$0.803\pm 0.001$	$0.792\pm 0.002$	$0.866\pm 0.002$	$\mathbf{0.149\pm 0.001}$	$\mathbf{0.253\pm 0.001}$	$\mathbf{2.556\pm 0.004}$	$\mathbf{0.000\pm 0.000}$	$\mathbf{0.000\pm 0.000}$
DR	$\mathbf{0.884\pm 0.001}$	$\mathbf{0.856\pm 0.002}$	$\mathbf{0.579\pm 0.002}$	$0.186\pm 0.001$	$0.280\pm 0.007$	$2.347\pm 0.028$	$\mathbf{0.000\pm 0.000}$	$\mathbf{0.000\pm 0.000}$

Appendix

Table 15: Residual refinement ablation on the MSL case study, including the unrepaired base model.

Method	$D_{\mathrm{o}}$ Centroid $\downarrow$	$D_{\mathrm{o}}$ Direction $\downarrow$	Q4 Acc. $\uparrow$	Q4 F1 $\uparrow$
Base	–	–	$0.222\pm 0.002$	$0.177\pm 0.010$
DR (w/o ref.)	$\mathbf{0.452\pm 0.080}$	$0.430\pm 0.013$	$0.222\pm 0.002$	$0.186\pm 0.011$
DR	$0.816\pm 0.062$	$\mathbf{0.278\pm 0.008}$	$\mathbf{0.476\pm 0.046}$	$\mathbf{0.238\pm 0.013}$

Appendix

Table 16: Representation and hard-example metrics for the MSL DR refinement ablation.

Dataset	Group	Class	Rank	ID	Acc.	NLL	ECE	$\mathbf{R_{c}}$
CIFAR-100	Easy	tractor	1	89	0.970	0.269	0.107	-4.959
CIFAR-100	Easy	orange	2	53	0.960	0.262	0.095	-4.922
CIFAR-100	Medium	table	50	84	0.860	0.637	0.111	-0.880
CIFAR-100	Medium	worm	51	99	0.820	0.611	0.104	-0.861
CIFAR-100	Hard	boy	99	11	0.530	1.519	0.200	8.952
CIFAR-100	Hard	girl	100	35	0.530	1.651	0.229	9.981

Appendix

Table 17: Selected obsolete classes for the large-label datasets. Classes are ranked by the base model’s reliability-risk score

R_{c}

, where lower values indicate easier/more reliable classes and higher values indicate harder/less reliable classes.

Panel of Figure 17 — Figure 17: Example images from the datasets used in our experimental evaluation.

Date	Season	Hour	Weather	Temp.	Humidity	$\ldots$	cnt
2011-01-01	1	0	1	0.24	0.81	$\ldots$	16
2011-01-01	1	1	1	0.22	0.80	$\ldots$	40
2011-01-01	1	2	1	0.22	0.80	$\ldots$	32

Appendix

Table 18: Example rows from the Bike Sharing hourly dataset.

Explore similar work

Apr 16, 2026cs.CV

Class Unlearning via Depth-Aware Removal of Forget-Specific Directions

Machine unlearning aims to remove targeted knowledge from a trained model without the cost of retraining from scratch. In class unlearning, however, reducing accuracy on forget classes does not necessarily imply true forgetting: forgotten information can remain encoded in internal representations, and apparent forgetting may arise from classifier-head suppression rather than representational removal. We show that existing class-unlearning methods often exhibit weak or negative selectivity, preserve forget-class structure in deep representations, or rely heavily on final-layer bias shifts. We then introduce DAMP (Depth-Aware Modulation by Projection), a one-shot, closed-form weight-surgery method that removes forget-specific directions from a pretrained network without gradient-based optimization. At each stage, DAMP computes class prototypes in the input space of the next learnable operator, extracts forget directions as residuals relative to retain-class prototypes, and applies a projection-based update to reduce downstream sensitivity to those directions. To preserve utility, DAMP uses a parameter-free depth-aware scaling rule derived from probe separability, applying smaller edits in early layers and larger edits in deeper layers. The method naturally extends to multi-class forgetting through low-rank subspace removal. Across MNIST, CIFAR-10, CIFAR-100, and Tiny ImageNet, and across convolutional and transformer architectures, DAMP more closely resembles the retraining gold standard than some of the prior methods, improving selective forgetting while better preserving retain-class performance and reducing residual forget-class structure in deep layers.

Arman Hatami, Romina Aalishah, Ilya E. Monosov

Johns Hopkins University

Exact Unlearning Unlearning Method Retraining Residual Connections Video Object Removal

Ask PDF

Sep 30, 2026cs.LG

Unmerge: Efficient Machine Unlearning via Task Arithmetic

Approximate machine unlearning seeks to remove the influence of a forget set from a trained model without full retraining. Existing gradient-based methods require data-dependent hyperparameter search, struggle when forget and retain knowledge are entangled, and offer little insight into where unlearning actually happens inside the network. We recast unlearning through the lens of task arithmetic: if finetuning produces a merged task vector

τ_m

that combines learning on forget and retain sets, unlearning is the inverse operation that subtracts a learned forget component

τ_F

to recover the retain task vector

τ_R

. The forget signal is concentrated: at every layer, forget activations lie in a subspace spanned by a handful of dominant directions, so we factorize

τ_F

in a low-rank forget basis, which is faithful up to a small tail-eigenvalue residual and limits how far the correction can perturb retain. We then optimize three intuitive goals (match the merged vector inside the forget span, suppress leakage into the retain span, and bound the correction size) that provably bound forget leakage and retain damage in activation space. The resulting algorithm, Unmerge, is fast and powerful: on class-level unlearning with ResNet-50 on CIFAR-100 and Tiny ImageNet, it improves Tug-of-War by up to ~24% over a baseline of comparable runtime and by up to ~18% over stronger baselines that run ~5x slower, keeps membership-inference exposure at the level of retraining, and shrinks the feature-distribution gap to the retrained model, where relabeling methods leave forget features cleanly separable. Further studies show that Unmerge also applies to ViT-S/16 and scales to Llama-3.2-3B. The per-layer basis geometry that drives the algorithm also serves as a layerwise diagnostic for when and where unlearning becomes structurally hard.

Haoran Tang, Andrew Tan, Rajiv Khanna

Department of Computer Science Purdue University

Machine Unlearning Retraining Merge Task Arithmetic Approximation

Ask PDF

Jul 27, 2026cs.LG

DECAF: De-Clustering for Adaptive Representational Unlearning

Machine unlearning, which aims to remove the influence of specific training data from a trained model, is a key requirement for privacy, accountability, and adaptive deployment. We argue that many unlearning methods are vulnerable to a simple clustering attack, which can recover class structure in an unsupervised manner, limiting their suitability for continual deployment where removal requests must be handled reliably on demand. To address this, we propose DECAF (DE-Clustering for Adaptive Forgetting), a post-hoc method that operates only on the forget set and is designed to break the cluster. DECAF combines input noise, confidence suppression, and entropy-based output diversification to disrupt the residual feature-space structure associated with forgotten data. On CIFAR-10 with ResNet-18, DECAF attains 0.10% forget-class accuracy, 79.4% retain accuracy, and an AUS of 0.88, surpassing all other baselines. In cluster-based analysis, it attains performance comparable to that of unlearning methods that use the full training set, while being significantly more efficient. Code: https://github.com/ale256/representation_unlearning.

Anjie Le, Can Peng, Hongcheng Guo +1

Institute of Biomedical Engineering, University of Oxford, UK · Fudan University.

Machine Unlearning Unsupervised Parameter-Efficient Adaptation Clustering

Ask PDF