PaperlayerView exact revision on arXiv

Source-linked AI summary

Semantic Concept Steering Breaks the Explanation Drift Loop in Continual Learning

Yehonatan Elisha, Oren Barkan, Noam Koenigstein

arXiv:2610.33877v1cs.CV

TL;DR

Continual learning can preserve accuracy while explanations drift toward spurious features, and pixel-level saliency regularization may reinforce that shift by using model-derived targets. C3L instead uses offline, class-consistent semantic concepts to guide relevance, and reports state-of-the-art results across five benchmarks and more than twenty baselines.

  • Problem

    Explanation-aware continual learning can preserve accuracy while attention shifts from diagnostic features to spurious correlations, a drift that standard metrics do not capture.

    Source: §1 · p. 2

  • Method

    C3L uses offline, model-state-independent, class-consistent semantic concepts to guide relevance and automates their discovery and spatial grounding.

    Source: §1 · pp. 2–3, §3.3 · p. 5

  • Results

    C3L outperforms more than twenty baselines across five benchmarks on accuracy, forgetting, semantic localization, spurious correlation mitigation, and OOD robustness.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The automated framework avoids manual annotation and is presented as a scalable form of concept-based supervision for continual learning.

    Source: §1 · p. 3, §5 · p. 10

    Crowding-out is a mechanism hypothesis rather than a guarantee because relevance conservation is approximate and the objective uses signed, min-max-normalized maps.

    Source: §3.3 · p. 6

Abstract

from arXiv · show

Catastrophic forgetting in continual learning (CL) manifests not only as accuracy degradation but also as \emph{explanation drift}: models maintain predictive accuracy while silently shifting attention from diagnostic features to spurious correlations. This work identifies and resolves a previously overlooked structural flaw in explanation-aware CL: pixel-level saliency regularization derives supervision targets from the model's own frozen saliency maps, creating a \emph{self-referential drift loop} that actively reinforces spurious attention across tasks. We introduce \textbf{C$^3$L} (Concept-Consistent Continual Learning), which breaks this loop by anchoring regularization to class-consistent semantic concepts (e.g., ``curved beak'', ``striped pattern'') discovered offline, independent of model state, providing (i) target independence from model drift, and (ii) cross-instance semantic consistency across every instance of a class. We identify that our concept-based objective, combined with approximate LRP relevance conservation, induces a \emph{crowding-out effect}: since concept and background regions compete for the relevance of the same prediction, enforcing high relevance within concept regions implicitly suppresses spurious background correlations without explicit penalization. \method does not require human-annotated concept sets, as it employs a fully automated concept discovery and spatial grounding. Extensive evaluation across five benchmarks and over 20 baselines demonstrates state-of-the-art performance in accuracy, forgetting, explanation quality, spurious correlation mitigation, and out-of-distribution robustness.

1 Introduction

Continual learning can preserve accuracy while explanations shift from diagnostic features to spurious correlations, a failure that standard metrics miss. C3L replaces model-derived saliency targets with class-consistent semantic concepts and reports strong results across five benchmarks.

1 Introduction

  • Models can retain predictive accuracy while attention shifts from diagnostic features to spurious correlations, a form of explanation drift not captured by standard metrics.If those correlations fail under distribution shift, performance can collapse.Source: §1 · p. 2
  • C3L anchors regularization to offline, model-state-independent semantic concepts shared across instances of a class.This provides target independence from model drift and cross-instance semantic consistency.Source: §1 · p. 2
  • Under approximate LRP relevance conservation, increasing relevance in concept regions can implicitly suppress background attributions through a crowding-out effect.A complementary stability term preserves background spatial structure while allowing its magnitude to shrink, steering freed relevance away from task-specific artifacts.Source: §1 · pp. 2–3
  • C3L automates concept discovery and spatial grounding, avoiding manual concept annotation for continual learning.The overview describes offline LLM discovery and SAM 3 grounding before continual training.Source: Figure 2 · p. 2, §1 · p. 3
  • C3L outperforms more than twenty baselines across five benchmarks on accuracy, forgetting, semantic localization, spurious correlation mitigation, and OOD robustness.The reported evaluation includes both CNN and ViT architectures.Source: §1 · p. 3

2 Related Work

Prior continual-learning approaches preserve predictions or model parameters but generally do not preserve semantic reasoning. C3L replaces model-dependent saliency targets with offline, class-consistent concept supervision in the continual-learning setting.

2 Related Work

  • Replay, regularization, architecture, prompt, and adapter methods address continual learning in parameter, feature, or sample space without preserving which visual features receive attention.The cited approaches include replay [5], regularization [6] [7], prompts [9] [22] [23], and adapters [24] [25] [27].Source: §2 · p. 3
  • Prior explanation-aware continual learning methods use pixel-level saliency, whose targets can reinforce already-drifted attention across tasks.RRR [10] stores Grad-CAM [42] maps; Saha and Roy and Liu et al. extend saliency-guided replay or masking.Source: §2 · pp. 3–4
  • Pixel-level attributions can be noisy, perturbation-sensitive, and prone to highlighting spurious correlations or high-frequency textures [59] [11].These concerns accompany the structural self-referential target problem.Source: §2 · p. 4
  • C3L uses offline LLM concept discovery [15] and SAM 3 spatial grounding [16] to generate dense semantic supervision without extensive manual annotation [56] [17].The framework uses these pipelines as preprocessing rather than constructing an inherently interpretable bottleneck model.Source: §2 · p. 4
  • C3L combines target independence, fine-grained semantic specificity, and class-shared cross-instance consistency, while stabilizing background structure rather than explicitly suppressing it.Table 13 isolates semantic specificity by comparing against object-level masks using the same SAM 3 model.Source: §2 · pp. 3–4

3 Method

C3L frames continual learning as preserving both classification knowledge and attention to class-defining concepts. It discovers and grounds class-consistent concepts independently of the learner, then uses concept and background regularization to address explanation drift.

3.1 Problem Formulation

  • C3L targets semantic stability in class-incremental learning: maintaining attention on meaningful class concepts in addition to maximizing accuracy.Training proceeds over sequential tasks with access to the current task's data and a small replay buffer.Source: §3.1 · p. 4

3.2 Automated Concept Discovery and Grounding

  • A label-free language-model approach generates class-discriminative, interpretable textual attributes, avoiding manual concept annotation.Concepts can optionally be filtered by their localization occurrence rate and mask coverage; unvalidated concept sets also achieve strong performance.Source: §3.2 · pp. 4–5
  • SAM 3 grounds the class concepts as masks for replay samples, with shared concept sets and learner-independent masks supporting cross-instance semantic consistency.Batched queries provide single-pass grounding without task-specific segmentation models or pixel-level annotations.Source: §3.2 · p. 5

3.3 Concept-Consistent Continual Learning

  • The concept-overlap objective enforces relevance within class-concept regions, using model-state-independent targets that cannot encode explanation drift.The framework describes its regularizers as a “correct-not-preserve” strategy rather than simply matching reference explanations that may contain spurious correlations.Source: §3.3 · p. 5
  • Under approximate LRP relevance conservation, increasing concept-region relevance can reduce background relevance without explicit background penalization, but this crowding-out mechanism is a hypothesis, not a guarantee.The authors note that conservation is approximate, relevance is signed, and the objective uses min-max normalized maps; they validate the mechanism empirically.Source: §3.3 · pp. 5–6
  • Background stability constrains the spatial pattern of background relevance while allowing its magnitude to decrease, and the authors report that semantic guidance remains effective without foundation models.The background term is scale-invariant and is intended to avoid drift toward task-specific artifacts while preserving contextual priors.Source: §3.3 · p. 6

4 Experiments

The experiments evaluate C3L across continual-learning benchmarks, architectures, and complementary tests of forgetting, explanation drift, component contributions, and out-of-distribution robustness. They report consistent gains over comparison methods, including better semantic localization and spurious-correlation mitigation.

4 Experiments

  • The experiments organize around standard continual-learning performance, explanation drift and spurious correlations, component ablations, and out-of-distribution robustness.The OOD evaluation focuses on ImageNet-A and ObjectNet; the explanation evaluation includes spurious-correlation accuracy and semantic localization.Source: §4 · p. 6

4.1 Experimental Setup

  • Evaluation spans five benchmarks with class-incremental task splits, ResNet-18 and ViT-B/16 architectures, and 20+ methods; reported metrics include Average Accuracy (ACC) and Average Forgetting (F).Results are averaged over 5 runs, and regularization is applied exclusively to replay samples with a buffer of 1,000.Source: §4.1 · pp. 6–7

4.2 RQ1: Standard Continual Learning Performance

  • C3L improves accuracy and forgetting across all datasets, outperforming state-of-the-art and explanation-based methods across ResNet-18 and ViT-B.Matched-explanation comparisons also favor concept-level over pixel-level regularization on both architectures.Source: §4.2 · p. 8, Table 1 · p. 8

4.3 RQ2: Explanation Drift and Spurious Correlations

  • C3L achieves the best performance across accuracy and semantic-localization metrics, and outperforms specialized debiasing methods on accuracy across both spurious-correlation datasets.The evaluation measures accuracy on FITZ and I-100C and semantic localization on CUB.Source: §4.3 · pp. 8–9, Table 2 · p. 9

4.4 RQ3: Ablation Studies

  • On ImageNet-R, removing either regularization term causes similar drops; on CUB, removing Lconcept causes larger drops, while removing Lstab also hurts.C3L remains superior to compared methods with a replay buffer of 500 samples, even methods using 1,000; with 200 samples it remains competitive and often matches or exceeds larger-buffer baselines.Source: §4.4 · pp. 9–10, Table 3 · p. 9, Table 4 · p. 10

4.5 RQ4: Robustness to Out-of-Distribution (OOD) Samples

  • C3L’s semantic regularization significantly increases accuracy on ImageNet-A and ObjectNet after fine-tuning, while RRR yields negligible gains or sometimes decreases accuracy.The comparison uses ResNet-18 (RRR/C3L) and ViT-B (RRR-ViT/C3L).Source: §4.5 · p. 10, Table 5 · p. 10

5 Conclusion

The conclusion presents C3L as a way to mitigate explanation drift by anchoring learning to semantic concepts, with automated supervision and reported state-of-the-art results across five benchmarks. The appendix is identified as the location for limitations and future work, alongside supplementary analyses and experiments.

5 Conclusion

  • C3L anchors representations to semantically grounded concepts to mitigate explanation drift and reports state-of-the-art continual-learning performance across five benchmarks, particularly under OOD shifts.The framework uses automated concept discovery and grounding rather than manual annotation.Source: §5 · p. 10
  • The authors direct readers to the appendix for limitations and future work.The appendix overview lists a dedicated limitations and future-work discussion.Source: §5 · p. 10, A Appendix Overview · p. 16

B Implementation Details

The implementation uses separate CNN and transformer backbones and explanation methods, with a shared Grad-CAM evaluation protocol. Training, replay, concept discovery, and spatial-grounding settings are specified for the experiments.

B Implementation Details

  • CNN experiments use ImageNet-1K-pretrained ResNet-18 with LRP ϵ-rule, while transformer experiments use ImageNet-21K-pretrained ViT-B/16 with AttnLRP.For unified saliency evaluation, baselines use Grad-CAM or Grad-CAM-ViT; C3L maps are min-max normalized to [0, 1] before BCE computation.Source: B Implementation Details · p. 16
  • Concept discovery follows Oikarinen et al. [15] with GPT-4o-mini, and SAM 3 spatial grounding uses confidence threshold 0.3.Concepts are filtered at occurrence rate ≥0.15 and coverage ≥0.2; prompts omit the “around” prompts from the original repository.Source: B Implementation Details · pp. 16–17
  • Regularization applies only to replay samples from a uniformly sampled buffer of 1,000; λconcept and λstab are 1.0 and 1.5 for CNNs, and 0.8 and 1.2 for ViTs.Hyperparameters were selected on validation data, and results average five runs with different seeds and task orderings; paired t-tests use p < 0.05.Source: B Implementation Details · pp. 16–17
  • Both architectures use SGD with batch size 128 and learning rate 0.001, training for 20 epochs on four datasets and 50 epochs on CUB.Baselines follow the setups recommended in their papers and code.Source: B Implementation Details · p. 16
  • CUB, ImageNet-R, and ImageNet-A use 10 tasks of 20 classes, while DomainNet uses five tasks of 69 classes.ObjectNet excludes three of its 113 classes before splitting the remaining 110 into 10 tasks of 11; methods share the same subsets and splits.Source: B Implementation Details · p. 16

C.1 Example Discovered Concepts

  • Table 6 gives selected examples of automatically discovered concepts, distinguishing concepts that passed filtering from those that did not.Bold concepts passed the filtering criteria; regular-text concepts were filtered out.Source: Table 6 · p. 17, §C.1 · p. 17

C.2 Concept Statistics Across Datasets

The paper summarizes automated concept generation and filtering across datasets, using occurrence and spatial coverage criteria to retain concepts.

C.2 Concept Statistics Across Datasets

  • Table 7 summarizes automated concept discovery and filtering across evaluated datasets, with averages calculated across concepts and classes.Source: Table 7 · p. 17, §C.2 · p. 17

C.3 Concept Validation Details

  • Occurrence rate is the fraction of a class’s images containing a given concept, measured by whether its segmentation set is nonempty.Source: §C.3 · p. 17
  • Spatial coverage is the average IoU between concept segments and class-specific object masks; concepts failing either the occurrence or coverage threshold are discarded.Object masks are created with SAM 3 using the class label as the prompt.Source: §C.3 · pp. 17–18

D Method Design Justification

C3L stabilizes background relevance rather than explicitly suppressing it, pairing that choice with concept-based relevance reallocation. The section also describes how relevance flow is measured across continual-learning tasks.

D.1 Background Regularization Strategy

  • Approximate relevance conservation lets high relevance in semantic concept regions crowd out background relevance, suppressing spurious cues without explicit penalization.Explicit suppression could remove valid diagnostic features omitted by automated concept discovery, while background can retain useful pretrained context such as pose, geometry, and scale.Source: §D.1 · p. 18, E Quantitative Evidence for the… · p. 18
  • Lstab preserves the spatial pattern of background relevance while its relative magnitude decreases, preventing drift toward new task-specific artifacts.Source: §D.1 · p. 18
  • With ViT-B on ImageNet-R and CUB, background stabilization outperforms both no background regularization and explicit background suppression without replay.The comparison with no background regularization isolates the background term; the no-replay alternative also removes replay.Source: §D.1 · p. 18

E.1 Relevance Mass Flow Analysis

  • The relevance-flow analysis tracks concept, background, and total relevance on the CUB T1 test set at tasks T1, T3, T5, T8, and T10.It measures each region’s share of the target logit and the concentration ratio, with approximate conservation predicting total relevance near 1.Source: §E.1 · p. 19

E.2 Results and Analysis

Across continual-learning tasks, relevance shifts from background regions toward semantic concepts, supporting the crowding-out account. Background stabilization preserves contextual patterns while the background relevance magnitude falls.

E.2 Results and Analysis

  • CR rose from 1.13 after T1 to 2.57 after T10 as concept relevance increased and background relevance decreased.Source: §E.2 · p. 19
  • A 0.20 gain in concept share matched a 0.18 reduction in background share, while total relevance remained 0.98–1.00 throughout training.Source: §E.2 · pp. 19–20
  • The authors attribute this reallocation to concept regularization under approximate relevance conservation, with Lstab preventing drift toward new spurious artifacts.Background relevance magnitude falls while its spatial pattern remains anchored to pretrained priors, preserving useful context.Source: §E.2 · p. 20
  • The analysis provides direct quantitative evidence for crowding-out by tracking relevance flow during training, rather than only measuring final explanation quality.Source: §E.2 · p. 20

F.1 Impact of Replay in Semantic-Centered Training

Full C3L achieves the best overall performance in the replay ablation, while C3L-NoReplay outperforms standard ER and performs comparably to or better than C3L-FullReplay on the reported datasets.

F.1 Impact of Replay in Semantic-Centered Training

  • Full C3L achieves the best overall performance, while C3L-NoReplay outperforms standard ER and surpasses C3L-FullReplay on CUB.C3L-NoReplay achieves comparable performance to C3L-FullReplay on ImageNet-R without storing past samples.Source: §F.1 · p. 21, Table 10 · p. 21

F.2 Explanation method

  • AttnLRP yields a clear advantage over Grad-CAM, IIA, and T-Attr, though C3L remains somewhat robust to explanation-method choice.The reported gap is relatively small when the attribution method produces reasonably faithful explanations.Source: §F.2 · p. 21, Table 11 · p. 21

F.3 Impact of Task Granularity

  • Across 5-, 10-, and 20-task ImageNet-R settings, C3L maintains the highest ACC and AIA and lowest F, with a significant margin over PLAN-R in the 20-task setting.Performance degrades for all methods as task count increases.Source: §F.3 · pp. 21–22, Table 12 · p. 22

F.4 Concept-Level vs. Object-Level Segmentation

  • Fine-grained concept masks consistently outperform object-level masks because they guide attention toward discriminative parts and attributes rather than whole objects.The comparison uses the same SAM 3 foundation model, isolating the contribution of concept-level guidance.Source: §F.4 · p. 22, Table 13 · p. 22

F.5 SAM 3 vs. Grounded SAM; F.6 Concept Filtering: Required vs. Optional

  • C3L’s performance degrades only slightly with GroundedSAM, and remains robust to unvalidated concepts, while filtering provides modest improvements.For verbose concepts, SAM 3 has an 86% detection hit-rate and 0.67 average confidence, versus GroundedSAM’s 59% and 0.45.Source: Table 14 · p. 22, §F.5 · p. 23, §F.6 · p. 23

F.7 SAM 3 Confidence Threshold; F.8 Hyperparameter Sensitivity Analysis

  • Performance varies smoothly across confidence thresholds and only slightly across tested hyperparameter ranges; τconf = 0.3 yields the best overall performance.The selected regularization weights are λconcept = 0.8 and λstab = 1.2, based on validation performance.Source: §F.7 · pp. 23–24, Table 16 · p. 23, §F.8 · p. 24

G.1 Performance on Non-Pretrained Networks; G.2 Full Continual Learning Results

  • C3L outperforms baselines by a large margin on CUB without pretraining; C3L++ is not the primary method because it does not use a continual-learning replay setting.The full-results table reports ACC, F, and AIA across five benchmarks, but the supplied passages do not provide its cell values.Source: §G.1 · p. 24, §G.2 · p. 24, Table 18 · p. 24

H.1 Explanation Quality

  • CUB localization gains are partly expected by design because C3L aligns relevance with object concept regions; FITZ and I-100C results provide complementary evidence independent of segmentation.The localization evaluation thresholds saliency maps at the 50th percentile and reports mIoU, pixAcc, and mAP.Source: §H.1 · pp. 24–26

H.2 Robustness Tests Further Details

C3L consistently outperforms explanation-aware methods in the reported out-of-distribution robustness evaluation. On PartImageNet, concept-guided learning also outperforms baselines without foundation-model priors.

H.2 Robustness Tests Further Details

  • C3L consistently outperforms explanation-aware methods in robustness to out-of-distribution data on ImageNet-A and ObjectNet.The evaluation fine-tunes on a 2,500-sample subset and measures accuracy and mIoU.Source: §H.2 · p. 26, Table 20 · p. 26

I.1 Performance without Foundation Models Usage

  • On PartImageNet, concept-guided C3L-PG outperforms all baselines when trained from scratch without foundation-model priors.The evaluation uses ground-truth part masks; equipping baselines with the same foundation-model guidance would be conceptually similar to applying C3L.Source: §I.1 · p. 27, Table 21 · p. 27

I.2 Isolating SAM 3’s Inherent Capabilities from C3L’s Gains

SAM 3’s standalone zero-shot accuracy falls substantially short of C3L, although the protocols are not directly comparable because zero-shot evaluation has no sequential training or forgetting dynamics. Controlled comparisons also favor fine-grained concept supervision over object masks.

I.2 Isolating SAM 3’s Inherent Capabilities from C3L’s Gains

  • SAM 3’s best zero-shot performance on CUB and ImageNet-R falls substantially short of C3L, but the protocols are not directly comparable because zero-shot evaluation lacks sequential training and forgetting dynamics.The comparison includes category-name and attribute-prompt variants using the same filtered concepts and SAM 3 configuration.Source: §I.2 · pp. 27–28, Table 22 · p. 28

I.3 Semantic Depth Over Object Segmentation

  • Using the same SAM 3 model, fine-grained concept-level supervision yields superior performance to object-level masks.The comparison supports semantic guidance, rather than merely foundation-model-derived segmentation, as a driver of results.Source: §I.3 · p. 28

I.4 Foundation Models as Automated Annotators

  • Foundation models act as offline automated annotators that provide binary spatial masks, not class predictions or latent representations to imitate.The masks specify where to attend, while the learner derives discriminative features from training labels.Source: §I.4 · p. 28

J Computational Analysis

C3L adds modest training-time and memory overhead, with preprocessing costs that can be managed or reduced. The qualitative examples show C3L retaining attention on semantic object regions, while the limitations include concept and segmentation failures and a vision-focused evaluation.

J Computational Analysis

  • On ImageNet-R, C3L increases end-to-end training time by 3% for ResNet-18 (28 min to 29 min) and 6% for ViT-B (1.5 h to 1.6 h).The added cost relative to RRR comes from LRP-based explanations and the two regularization terms.Source: J Computational Analysis · p. 29
  • Preprocessing takes 63 minutes total, or 38 minutes without optional concept validation; asynchronous processing or smaller replay buffers can reduce deployment burden.C3L with |M| = 200 remains competitive and outperforms baselines using |M| = 1000; C3L-NoReplay eliminates replay preprocessing.Source: J Computational Analysis · p. 29
  • C3L adds approximately 8% memory usage over RRR, primarily for binary concept masks.For 1,000 replay samples, C3L stores 12 MB of concept masks and 23 MB of reference saliency maps, while RRR stores 26 MB of Grad-CAM maps.Source: J Computational Analysis · p. 29

K Qualitative Visualizations

  • After continual learning, C3L preserves attention on semantically meaningful object regions, while baseline explanations shift toward backgrounds and texture cues.The qualitative examples use an ImageNet-pretrained ViT-B trained on ObjectNet.Source: K Qualitative Visualizations · p. 29, Figure 3 · p. 30

L.1 Concept Discovery Failures

  • Concept discovery can fail for abstract, rare, or ambiguous attributes that lack reliable visual grounding, occur in fewer than 15% of instances, or lack clear spatial boundaries.Examples include “natural habitat,” “open beak,” and “dark color.”Source: §L.1 · pp. 29–30

L.2 Segmentation Failures

  • SAM 3 may miss small or occluded features covering less than 2% of image area, and polysemous concepts can produce occasional false-positive masks.The small-feature issue primarily affects fine-grained datasets such as CUB.Source: §L.2 · p. 30

M Limitations and Future Work

  • C3L depends on foundation models, adds preprocessing cost, and is evaluated only on vision tasks; its model-derived background reference attenuates existing spurious cues rather than fully removing them.The authors note that extending the approach to language, audio, or multimodal settings requires domain-specific concept discovery.Source: M Limitations and Future Work · p. 30
Loading 2610.33877v1…