Out-of-Distribution Detection in Continual Learning

AI summary

Overview

Research area: Machine learning — continual (class-incremental) learning combined with out-of-distribution (OOD) detection, i.e. the "open-world" problem of a model that keeps learning new tasks while still recognizing when an input belongs to none of the classes it has learned.

Technical level: Advanced. The paper assumes familiarity with class-incremental learning (CIL), prompt-based continual learning, Kolmogorov–Arnold Networks (KANs), radial basis functions, and standard OOD scoring functions (MSP, energy, MaxLogit, kNN-based scores).

Scope: The paper proposes a rehearsal-free framework for OOD detection in continual learning, built on a task-adaptive RBF-KAN classifier (TA-KAN) plus a prompt-space geometry guidance score, and evaluates it on ImageNet-R, CUB-200 and CIFAR-100 against several classifier variants and three OOD-CL baselines.

What This Paper Is About

A model that learns tasks one after another must do two things at once: remember earlier tasks, and notice when a test input belongs to none of the classes it has seen so far. The second job is hard in continual learning because the model no longer has the old training data, and the usual confidence-based OOD scores become unreliable — they are especially poor at flagging OOD inputs as unknown when those inputs sit near older classes. The paper's goal is to fix this bias without storing any past data, using a KAN-based classifier whose activation locality is tuned per task, combined with a geometric score computed in prompt space rather than feature space.

Key Contributions

1. A rehearsal-free OOD-CL framework that targets practical constraints the authors identify — memory, privacy, and reliability — instead of relying on a buffer of past exemplars or a closed-world assumption.
2. TA-KAN, a task-adaptive KAN classifier for RBF-KAN-based heads: each task gets its own RBF width factor, chosen from a deterministic schedule based on task recency, so older tasks get sharper (narrower) activations. It keeps a single shared RBF layer and adds no learnable parameters, preserving the parsimony required in CIL.
3. A geometry-guided OOD score that multiplies classifier confidence by a similarity-to-nearest-prompt term computed over learned task keys. It needs no past in-distribution data and operates in prompt space, which the authors describe as encoding higher-level task semantics than the embedding space.
4. A two-level manifold interpretation: the combined score is written as (class alignment) × (task alignment), so a sample is treated as in-distribution only if it aligns with both a learned class prototype and a learned task prototype.

Main Findings

• KAN improves OOD detection only modestly relative to its ID gains. Replacing the linear head with a KAN head (KAN-Head) improves CIL performance and ID accuracy, but the OOD improvement over a linear head is limited compared with the ID improvement (Figure 2).
• KAN heads are biased toward recently learned classes. In a trained CIL model on ImageNet-R (5-task), evaluated with each task in turn treated as ID, KAN-Head shows a much higher false positive rate when older tasks are the ID set than when recently learned tasks are, meaning OOD samples are frequently confused with old ID classes (Figure 1).
• TA-KAN reduces that recency bias. On CUB-200 with 5 tasks, TA-KAN reaches FPR_L 52.11 and AUC_L 85.86 versus KAN-Head's 59.65 and 83.48, while ACC_L is 80.86 versus 81.55. With 20 tasks, TA-KAN reaches FPR_L 51.35 and AUC_L 84.42 versus KAN-Head's 59.69 and 81.86 (ACC_L 75.70 vs 75.24).
• Guidance adds further OOD gains. TA-KAN + Guidance on CUB-200 with 5 tasks gives FPR_L 48.73, AUC_L 86.42, AUP_In 85.59, AUP_Out 86.96 at unchanged ACC_L 80.86. The 10-task setting gives FPR_L 68.28 and AUC_L 81.99 (versus 70.63 and 81.55 for TA-KAN alone); the 20-task setting gives FPR_L 50.48, AUC_L 84.77, AUP_In 84.85, AUP_Out 85.10.
• ImageNet-R shows the same pattern with smaller margins. TA-KAN attains FPR_L 40.07 / AUC_L 85.77 / ACC_L 78.83 at 5 tasks, 40.25 / 86.76 / 77.76 at 10 tasks, 42.57 / 85.99 / 77.03 at 20 tasks, and 44.16 / 86.48 / 72.38 at 40 tasks. Adding guidance gives 39.72 / 85.92, 39.98 / 86.87, 42.61 / 86.10, and 44.03 / 86.41 with unchanged ACC_L.
• Baseline comparison on ImageNet-R (5 tasks): Linear-Head 42.74 FPR_L / 85.44 AUC_L / 77.96 ACC_L; MSP-BC-CE 40.65 / 85.50 / 79.25; KAN-Head 42.52 / 84.81 / 79.25; TA-KAN + Guidance 39.72 / 85.92 / 78.83.
• Guidance leaves ID accuracy untouched in the reported tables (identical ACC_L for TA-KAN and TA-KAN + Guidance), consistent with it being an OOD-only rescoring of classifier confidence.
• CIFAR-100 results are not reported in the supplied content. The paper lists CIFAR-100 among its three benchmarks, but the text is truncated before those results; comparisons with standard OOD methods (MSP, MaxLogit, Energy, Entropy, KL) and ablation details referenced to Section 4.3 and Appendix E.3 are also outside the provided excerpt.

Methodology in Plain English

The authors start from CODAPrompt, a prompt-based continual learner that keeps a frozen backbone and attaches small learnable prompts plus keys per task; a query–key cosine similarity decides which prompts a sample should use. They swap the usual linear classifier for an RBF-KAN head, where each feature dimension is expanded over a set of Gaussian bumps centred on a fixed grid, and the bump width is normally uniform.

Their key observation is that width controls locality: narrow bumps give sharp, local responses, wide bumps give smooth, general responses. So they give each task a different width factor α_t, set by a simple linear schedule from α_min for the oldest task up to α_max = 1 for the newest (0 < α_min ≤ α_t ≤ α_max ≤ 1). No extra parameters are learned. At training time the true task identity picks the width; at test time the width is picked by the query–key match, which for an OOD input points to the task region where the model is most likely to be confused. Older, harder tasks therefore get sharper decision regions while newer tasks keep the generalisation they need.

Because a continual learner is optimised to generalise (which lowers confidence on genuinely in-distribution data), confidence alone is not enough. The authors multiply the standard confidence score (MSP by default) by a guidance term: the maximum cosine similarity between the sample's query and any learned task key, rescaled from [-1, 1] to [0, 1]. Empirically they found taking the single most similar prompt (k = 1) slightly outperformed averaging over all M prompts per task.

Evaluation follows standard CIL protocol: before learning each new task, test data from all previously learned tasks are treated as in-distribution and test data from the incoming task as out-of-distribution, giving N−1 detection rounds. Metrics are AUC_L, FPR_L, AUP_L_In, AUP_L_Out and ACC_L, plus their incremental counterparts. All experiments use a non-exemplar setting with randomly shuffled task orders across seeds.

Why This Matters

Impact on research. The paper reframes OOD detection in continual learning as a bias problem rather than purely a scoring problem: the classifier's decision regions themselves drift toward new classes as tasks accumulate, and no post-hoc score can fully repair that. It also introduces KAN architectures into the open-world continual setting, where they had mainly been studied under closed-world assumptions, and shows the transfer is not automatic.

Real-world applications:

• On-device personalization where user data cannot be stored or replayed (keyboards, photo apps, wearable health monitors), so rehearsal buffers are not an option.
• Medical imaging systems that are extended with new diagnostic categories over time and must flag scans outside all known categories for human review.
• Autonomous perception and robotics, where new environments or object classes are added incrementally and unknown inputs must trigger caution rather than a confident wrong answer.
• Security and fraud monitoring, where new attack patterns appear continuously and the model must recognize "none of the above" without retaining sensitive historical records.

Industry relevance. The no-buffer constraint directly matches privacy regulation and memory budgets on deployed devices, and the classifier swap is described as low-cost: the same RBF layer is shared across all tasks, and the width factors are a predefined schedule rather than learned parameters. The paper also states the approach generalizes to prompt-based models beyond CODAPrompt (Appendix E.5), which matters for teams already running prompt-tuning pipelines.

Future Directions

• Fuller benchmark coverage. CIFAR-100 is named as a benchmark but its results are absent from the provided content; completing that comparison, including against the standard OOD scores (MSP, MaxLogit, Energy, Entropy, KL), is the obvious next validation step.
• Learning the width schedule instead of fixing it. The α_t values come from a linear, deterministic schedule between α_min and α_max; whether a learned or data-driven schedule does better, without breaking parsimony, is open.
• Testing on other prompt-based and larger pre-trained backbones. The authors point to Appendix E.5 for generalization beyond CODAPrompt, but broader evidence across prompt methods and foundation models would be needed before deployment claims.
• Longer and more open-ended task streams. The protocol treats future-task classes as OOD until they are learned, and the paper notes these samples eventually become in-distribution; how the score behaves under much longer streams, and how the decision threshold should be set without past ID data, remain practical questions.

Target Audience

Researchers and graduate students working on continual learning, OOD detection, or both, who already know the standard scoring functions and CIL protocols. It is also useful for applied ML engineers building models that must be updated with new classes on devices without data retention, though they should expect to read the appendices for implementation and ablation details, which are outside the excerpt summarized here.

Authors’ abstract

Out-of-distribution detection in continual learning (OOD-CL) is highly challenging, as it requires (i) learning a sequence of tasks without catastrophic forgetting, and (ii) identifying unknown samples (OOD) without access to known (in-distribution (ID)) data from past tasks, where standard confidence-based OOD scores tend to degrade. Existing methods largely rely on a rehearsal buffer of past data or assume closed-world settings, limiting their scalability and reliability in real-world scenarios. In this paper, we propose a novel OOD detection framework for rehearsal-free continual learning, inspired by Kolmogorov-Arnold Networks (KAN), which mitigate catastrophic forgetting via their inherent local neuroplasticity. To better leverage KAN in OOD-CL, we introduce a task-adaptive classifier architecture, TA-KAN, which mitigates recency bias and enhances ID--OOD separability through task-specific control of activation locality. Furthermore, we propose a geometry-guided OOD score that complements TA-KAN classifier confidence without relying on past ID data. Our method significantly improves OOD detection while maintaining strong ID classification performance, achieving state-of-the-art results in OOD-CL.