Semantic Concept Steering Breaks the Explanation Drift Loop in Continual Learning

AI summary

Overview

Research area: Computer Vision — continual learning (class-incremental), explainable AI (saliency/attribution methods), and concept-based interpretability. Technical level: Intermediate to Advanced (requires familiarity with continual learning benchmarks, LRP/AttnLRP relevance propagation, and vision-language foundation models). Scope: The paper diagnoses a structural flaw in explanation-aware continual learning — a "self-referential drift loop" — and proposes C³L, a concept-anchored regularization framework that breaks it, evaluated on five benchmarks against more than 20 baselines on ResNet-18 and ViT-B.

What This Paper Is About

Continual learning models can keep their accuracy while silently changing why they make predictions, shifting attention from diagnostic features (a bird's beak, a guitar's fretboard) to spurious background cues. This "explanation drift" is currently fought with pixel-level saliency regularization, but those methods take their supervision targets from the model's own frozen saliency maps — so if the model has already drifted, the drift becomes the training target. The paper's goal is to break that loop by anchoring regularization to class-consistent semantic concepts that are discovered offline, independently of the learner.

Key Contributions

1. Problem identification: The authors formalize the self-referential drift loop as a structural property of any explanation-aware continual learning scheme that derives supervision targets from model-state-dependent saliency.
2. Crowding-out mechanism: They identify that under approximate LRP relevance conservation, concept and background regions compete for the same prediction's relevance, so enforcing relevance inside concept regions implicitly suppresses spurious background attention without explicit penalization — and they validate this quantitatively.
3. The C³L framework: A concept-guided continual learning method combining a concept overlap term with a scale-invariant background stability term, requiring no human concept annotation (LLM-based concept discovery plus SAM 3 spatial grounding), applicable to both CNNs and ViTs.
4. State-of-the-art results: C³L outperforms more than twenty baselines across five benchmarks on accuracy, forgetting, semantic localization, spurious correlation mitigation, and out-of-distribution robustness.

Main Findings

• Accuracy and forgetting gains: On ResNet-18, C³L reaches CUB ACC 79.73 ± 0.31 with forgetting 5.32 ± 0.63, versus the strongest explanation-aware baseline RRR at 72.21 ± 0.78 with forgetting 9.10 ± 0.44. On ViT-B it reaches CUB ACC 83.65 ± 0.28 (forgetting 2.92 ± 0.24), versus PLAN-R at 79.24 ± 0.56 (forgetting 3.52 ± 0.28).
• Consistency across benchmarks: ResNet-18 results are DomainNet 73.01, ImageNet-R 74.35, ImageNet-A 67.84, ObjectNet 71.12 (ACC); ViT-B results are DomainNet 80.43, ImageNet-R 82.74, ImageNet-A 71.14, ObjectNet 76.83 (ACC). Full result tables including all methods are deferred to Table 19 in the appendix.
• Gains are not just from better saliency: Controlled variants RRR-LRP and RRR-ViT-AttnLRP swap Grad-CAM for the paper's explanation methods. RRR-ViT-AttnLRP improves over RRR-ViT (CUB 78.09 vs 76.94), but C³L using the same AttnLRP improves far more — evidence that concept-level targets beat pixel-level targets, not that LRP alone wins.
• Not just foundation-model capability: Removing foundation models entirely (training on PartImageNet with ground-truth annotations) still leaves C³L ahead of all baselines; SAM 3's standalone zero-shot classification falls well short; and fine-grained concept masks outperform coarse object-level masks from the same SAM 3 model.
• Crowding-out confirmed empirically: The authors treat crowding-out as a mechanism hypothesis rather than a guarantee, since conservation is only approximate, relevance is signed, and the loss operates on min-max normalized maps; it is validated in Appendix E.
• Background stabilization beats background suppression: Table 8 reports that the stability design outperforms removing background regularization; the cosine-similarity formulation constrains where background relevance lies while letting its magnitude shrink.
• Robustness to imperfect concepts: Appendix F.6 shows concept filtering is optional — unvalidated concept sets still perform strongly — and Appendix F.5 reports robustness under weak concept grounding.
• Modest cost: About a 3% training-time increase relative to RRR on ResNet-18, and only 8% additional storage versus RRR; with a 200-example buffer the method is still competitive with state-of-the-art baselines using 1000.
• Smaller buffer works: Section 4.4 reports C³L with |M| = 200 is competitive against all baselines run at |M| = 1000.
• Not reported in the provided content: The specific numeric outcomes for the spurious-correlation (Fitzpatrick17K, Split ImageNet-100C) and semantic localization (mIoU, pixAcc, mAP) evaluations in Sections 4.3–4.5, and the ablation numbers in Section 4.4, are referenced but their values are not included in the text supplied.

Methodology in Plain English

The setup is standard class-incremental learning: a model sees a sequence of tasks with disjoint classes, keeps a small replay buffer (|M| = 1000 by default), and must classify across all classes seen so far.

The pipeline has three moving parts:

• Build concept targets offline, before training. For each class, an LLM is queried to produce discriminative textual attributes (e.g., "curved beak," "striped pattern"), following the label-free concept approach of Oikarinen et al. Optionally, concepts are filtered on held-out data by occurrence rate ν_occ ≥ τ_occ and coverage ν_cov (IoU against the class-object segment); failing concepts are pruned. Then SAM 3 turns each concept phrase into binary masks on the replay images, keeping only masks with confidence s_i ≥ τ_conf; their union is the concept mask M_concept, and its complement is the background mask. Because these masks are made without looking at the learner, they cannot encode its drift, and because concept sets are shared across all instances of a class, every instance gets the same semantic supervision.
• Anchor relevance to concepts. During training on each task, the total loss is the task cross-entropy plus λ_concept·L_concept plus λ_stab·L_stab. Explanations come from LRP with the ε-rule (ε = 10⁻⁶) for CNNs and AttnLRP for ViTs. L_concept applies a binary cross-entropy between the saliency restricted to concept regions and the concept mask itself — a "correct-not-preserve" strategy rather than matching a reference saliency map.
• Let crowding-out do the suppression, and only stabilize the background. Under approximate relevance conservation, the relevance inside and outside concept regions must sum to the target logit, so pushing concept relevance up pushes background relevance down. L_stab stores a reference saliency when a sample enters the buffer and penalizes cosine drift of the background region in later tasks — scale-invariant, so background magnitude stays free to shrink. The authors argue this does not reintroduce the drift loop because the corrective pressure comes from the model-independent concept term.

Evaluation uses ACC, forgetting F, and AIA, averaged over 5 runs with paired t-tests (p < 0.05). Architectures are ResNet-18 (ImageNet-1K pretrained) and ViT-B/16 (ImageNet-21K pretrained). Datasets are split into 10 sequential tasks (5 for DomainNet); for ObjectNet, 3 randomly sampled classes are excluded so the remaining 110 divide evenly.

Why This Matters

Explanation drift is invisible to standard continual learning metrics but matters practically: when a model has quietly learned to rely on background cues, that shortcut can collapse under distribution shift. This paper reframes the fix as a supervision-source problem rather than a loss-function problem, which is a cleaner and more general framing than prior saliency-alignment work.

Real-world applications:

• Medical imaging systems that learn new diseases sequentially, where attention must stay on lesions rather than scanner-specific artifacts.
• Autonomous driving adapting to new environments, where decisions must rest on road-relevant features rather than weather or lighting shortcuts.
• Industrial inspection and remote sensing, where sequential deployment across product lines or regions invites spurious background correlations.
• Regulated AI deployments (finance, hiring, diagnostics) where auditors need evidence that the model's reasoning, not just its accuracy, is stable over time.

Industry relevance: the method needs no manual concept annotation and adds only ~3% training time and ~8% storage over RRR, which is a realistic retrofit for teams already running replay-based continual learning. Concept masks are class-shared and computed offline, so the supervision pipeline is cached and reusable across training runs.

Future Directions

• Extending to concept bottleneck models. The authors explicitly state that systematically adapting CBMs to continual learning is a separate research problem and leave it to future work.
• Tightening the crowding-out claim. Because conservation is approximate, relevance is signed, and the loss works on normalized maps, crowding-out is presented as a hypothesis validated empirically rather than a theorem; a formal characterization is an open question.
• Reducing the preprocessing cost. The offline LLM-plus-SAM 3 pipeline adds a preprocessing stage; the authors note this can be substantially reduced by using a smaller replay buffer but do not fully eliminate it.
• Testing whether concept targets transfer across domains and task granularities. The paper ablates task granularity, spatial grounding, and concept discovery mechanism (RQ3), but the summary content does not report those values, leaving the boundaries of concept quality — and how much concept noise can be tolerated — as an open practical question.

Target Audience

Researchers and graduate students working on continual learning, explainable AI, or concept-based interpretability; practitioners deploying sequential-training systems that must stay auditable; and engineers interested in composing LLM-based concept discovery with open-vocabulary segmentation as an offline supervision source. Readers should be comfortable with replay-based continual learning, attribution methods (LRP, Grad-CAM), and standard CL metrics before diving in.

Authors’ abstract

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.