Classifier-Centric Adaptive Framework

• Classifier-centric adaptive framework is a design principle that concentrates adaptation on classifier modules rather than full end-to-end retraining.
• It employs techniques like classifier decomposition, parameter fusion on frozen backbones, and specialist routing to handle bias, domain shifts, and efficiency.
• Empirical results in settings such as imbalanced learning, continual learning, and fairness demonstrate improved accuracy and robustness with targeted classifier updates.

Searching arXiv for the cited papers and closely related work on classifier-centric adaptive frameworks. Classifier-centric adaptive framework denotes a family of learning designs in which the primary locus of adaptation is the classifier itself, classifier-adjacent modules, or classifier selection logic, rather than indiscriminate end-to-end retraining. Across the cited literature, the phrase is used for methods that split the classifier into bias-handling and unbiased components, fuse task-specific classifier parameters, route inputs to the most competent classifier, condition classifier layers on demographic groups, align domains with class-specific domain classifiers, or learn class-wise conformal penalties. In that sense, the term refers less to a single canonical architecture than to a recurring design principle: adaptation is concentrated where decisions are made, so that stability, fairness, robustness, or efficiency can be improved without treating the entire representation stack as the only object of optimization (Wang et al., 2022, Guo et al., 25 Mar 2025, Bouazzaoui et al., 2023, Gong et al., 2020, Tang et al., 2020, Marani et al., 14 Jan 2026).

1. Conceptual scope and defining characteristics

A classifier-centric adaptive framework is classifier-centric because it makes the classifier the place where a task-specific pathology is absorbed or controlled. In imbalanced semi-supervised learning, the central claim is that imbalance bias should be handled in the classifier rather than only by repairing pseudo-labels or re-balancing data; the proposed solution is a bias adaptive classifier composed of a bias attractor and the original linear classifier (Wang et al., 2022). In class-incremental learning, the emphasis is on preserving classifier discriminability over a growing label space by freezing CLIP and adapting only a small parameter module whose parameters are fused across tasks (Guo et al., 25 Mar 2025). In FPGA deployment, the framework is classifier-centric because the adaptive mechanism decides which classifier to use for each incoming sample instead of changing model weights online (Bouazzaoui et al., 2023).

The same logic appears in other settings. Open-vocabulary camouflaged object segmentation treats classification quality as the main lever for improving segmentation and adapts a lightweight text classifier while keeping the CLIP backbone frozen (Zhang et al., 29 Sep 2025). Group adaptive face recognition mitigates bias through demographic-conditioned classifier layers and an automation module that decides where specialization is needed (Gong et al., 2020). Unsupervised adaptive object detection replaces a single domain discriminator with a domain classifier bank, one classifier per category plus background, so that alignment is class-conditioned rather than generic (Tang et al., 2020). Class Adaptive Conformal Training moves from one global conformal penalty to one multiplier per class, thereby adapting the classifier to class-conditional efficiency constraints (Marani et al., 14 Jan 2026). In strategic classification, the classifier is explicitly optimized as a mechanism that shapes incentives for constructive feature changes (Chen et al., 2020). In Adaptive Weighted Deep Forest, each cascade level updates instance weights according to classifier output quality, making later levels focus more strongly on hard instances (Utkin et al., 2019).

Setting	Adaptive locus	Representative paper
Imbalanced SSL	Bias attractor + linear classifier	(Wang et al., 2022)
Class-incremental learning	Adapter fusion on frozen CLIP	(Guo et al., 25 Mar 2025)
FPGA inference	Dynamic classifier selection	(Bouazzaoui et al., 2023)
Face recognition fairness	Group adaptive kernels and attention	(Gong et al., 2020)
UDA object detection	Domain classifier bank	(Tang et al., 2020)
Conformal prediction	Class-wise multipliers and penalties	(Marani et al., 14 Jan 2026)

A common misconception is that classifier-centric methods only modify the final linear layer. The surveyed work uses the term more broadly. The adaptive component may be a residual bias module, a low-rank parameter fusion rule, a text adapter, a bank of class-level domain classifiers, a demographic-conditioned convolution block, or a runtime selection mechanism (Wang et al., 2022, Guo et al., 25 Mar 2025, Zhang et al., 29 Sep 2025, Gong et al., 2020, Tang et al., 2020, Bouazzaoui et al., 2023).

2. Recurrent architectural patterns

One recurrent pattern is classifier decomposition. In the bias adaptive classifier for imbalanced semi-supervised learning, the classifier is written as

$$F_{\omega,\phi}(\mathbf z)=(\mathbf I+\Delta f_\omega)\circ f_{\phi}^{\rm cls}(\mathbf z),$$

where $f_{\phi}^{\rm cls}$ is the standard linear classifier and $\Delta f_\omega$ is the bias attractor. The attractor is a lightweight MLP with one hidden layer attached through a residual connection, so it can absorb training bias during optimization and then be removed at test time, leaving the linear classifier in a cleaner state (Wang et al., 2022).

A second pattern is parameter-efficient classifier adaptation on frozen pretrained backbones. In class-incremental learning with CLIP, the encoder and text side remain fixed while an adaptive parameter module containing a four-layer linear transformation with attention-based feature enhancement is updated, and its output is fused with the original CLIP image feature: $$z'_{t-1} = (1-\lambda)\,\phi_{fc}(z_{t-1}) + \lambda z_{t-1}.$$ This preserves part of CLIP’s original representation while allowing task-specific adaptation (Guo et al., 25 Mar 2025). Open-vocabulary camouflaged object segmentation follows a closely related template, but on the text side: a lightweight adapter at the final textual layer transforms $\mathbf z$ by $A^t(\mathbf z)$ and applies the residual update $\mathbf y=\mathbf z+s\cdot A^t(\mathbf z)$, with only about $0.18$M trainable parameters (Zhang et al., 29 Sep 2025).

A third pattern is specialist banks and routing mechanisms. The domain classifier bank in adaptive object detection is

$\mathcal{D} = \{\mathcal{D}_i\}_{i=1}^{C+1},$

with one domain classifier per object class plus background, so that region features are aligned by category rather than through a single instance-level discriminator (Tang et al., 2020). The FPGA real-time adaptive neural network uses an ensemble of five neural network models, a competence estimator, and partial reconfiguration; only the selected model is used for inference, and the adaptive behavior lies in dynamic selection rather than in parameter updates (Bouazzaoui et al., 2023).

A fourth pattern is conditional classifier specialization. Group Adaptive Classifier forms demographic-specific kernels by masking shared kernels with group-specific masks and modulates channels with group-specific attention, while an automation module decides which layers should remain adaptive by measuring the dissimilarity among demographic-adaptive parameters (Gong et al., 2020). This suggests a broader interpretation of classifier-centricity: the classifier may remain a single network, but parts of it are explicitly conditioned on the subgroup relevant to the current sample.

3. Adaptive signals and optimization mechanisms

Classifier-centric adaptation is distinguished not only by where adaptation occurs but by how it is controlled. In class-incremental learning with CLIP, the central novelty is stacking-based adaptive weighted parameter fusion. If adjacent task adapters are $W_{t-1}=B_{t-1}A_{t-1}$ and $f_{\phi}^{\rm cls}$0, the method stacks the factors across tasks and introduces a balance factor $f_{\phi}^{\rm cls}$1: $f_{\phi}^{\rm cls}$2 The value of $f_{\phi}^{\rm cls}$3 is computed from the distribution relationship between adjacent tasks using MMD and LDA, and

$f_{\phi}^{\rm cls}$4

The stated purpose is to balance adjacent-task distribution alignment and distinguishability (Guo et al., 25 Mar 2025).

In imbalanced semi-supervised learning, the adaptive mechanism is bi-level rather than purely feed-forward. The lower level fits labeled and pseudo-labeled data with the modified classifier, while the upper level updates the bias attractor on a class-balanced labeled batch. The lower-level update is

$f_{\phi}^{\rm cls}$5

and the upper-level update is

$f_{\phi}^{\rm cls}$6

The paper’s proposition further rewrites the update of $f_{\phi}^{\rm cls}$7 in a form where $f_{\phi}^{\rm cls}$8 measures the similarity between a sample’s gradient and the average gradient on the balanced set, making the classifier adapt according to whether the sample’s influence aligns with balanced training behavior (Wang et al., 2022).

In conformal prediction, the adaptive signal is class-wise constraint violation. Class Adaptive Conformal Training replaces one global regularization weight with per-class multipliers $f_{\phi}^{\rm cls}$9 and penalties $\Delta f_\omega$0, optimized through an augmented Lagrangian method: $\Delta f_\omega$1 The multiplier update is class-specific, and $\Delta f_\omega$2 is increased when the class-wise set-size constraint does not improve (Marani et al., 14 Jan 2026).

Other frameworks use different supervisory signals but preserve the same principle. The domain classifier bank crosses teacher confidence with domain-classifier entropy to modulate class-conditioned adversarial alignment and mean-teacher consistency (Tang et al., 2020). Strategic classification optimizes

$\Delta f_\omega$3

so that the classifier discourages manipulation while encouraging improving best responses (Chen et al., 2020). Adaptive Weighted Deep Forest computes instance weights from the distance between a mean class vector and the one-hot target vector, so that later cascade levels emphasize difficult examples rather than removing easy ones through a hard confidence threshold (Utkin et al., 2019).

4. Application domains and task-specific interpretations

The framework has been instantiated in several technically distinct problem classes. In continual learning, the classifier-centric interpretation is tied to stability and plasticity. The CLIP-based class-incremental method treats incremental learning as a problem of evolving the classifier-adapter parameters without destroying prior decision structure, while ACL is described as compatible with prompt tuning, adapter tuning, and classifier-centric methods because it inserts a short adaptation phase before the core continual-learning phase for each task (Guo et al., 25 Mar 2025, Lu et al., 4 Jun 2025).

In semi-supervised and long-tailed recognition, the classifier is the location where imbalance is assimilated rather than merely corrected downstream. The bias adaptive classifier argues that an initially biased classifier tends to assign minority-class unlabeled samples to majority classes, creating confirmation bias; its decomposition explicitly separates fitting the biased training stream from preserving an unbiased linear decision boundary (Wang et al., 2022). CaCT addresses a related asymmetry in uncertainty quantification by allowing different classes to incur different conformal penalties, particularly when class difficulties differ substantially or the data are long-tailed (Marani et al., 14 Jan 2026).

In perception systems, classifier-centric adaptation often appears as a way to inject semantics into a more complex pipeline. Open-vocabulary camouflaged object segmentation argues that classification is not just auxiliary but a major determinant of segmentation quality, so a lightweight text adapter and layered asymmetric initialization are used to strengthen semantic recognition under camouflage (Zhang et al., 29 Sep 2025). Unsupervised adaptive object detection makes the same move in domain adaptation: instead of aligning instances generically, it aligns region features conditioned on object category through a class-specific domain classifier bank (Tang et al., 2020).

In fairness and deployment-time interaction, the classifier becomes a control surface for social or behavioral effects. Group Adaptive Classifier conditions convolution kernels and channel-wise attention on demographic attributes and couples this with a de-biasing loss on intra-class spread across groups (Gong et al., 2020). Strategic classification frames prediction and adaptation as a Stackelberg game in which the classifier is designed to shape how decision subjects alter improvable and manipulable features (Chen et al., 2020). FPGA-based dynamic classifier selection further broadens the category: here the classifier-centric adaptive mechanism is neither fairness-oriented nor statistical in the usual sense, but a sample-specific hardware-aware routing policy over five neural network models (Bouazzaoui et al., 2023).

5. Empirical behavior across the literature

The empirical record reported in these works is consistently framed around the claim that concentrating adaptation in the classifier can improve a task-specific trade-off. In class-incremental learning with CLIP, the method achieves the best or near-best average incremental accuracy and strong last-task accuracy on ImageNet100 and CIFAR100, reaching about $\Delta f_\omega$4 average accuracy and around $\Delta f_\omega$5 last-task accuracy on ImageNet100, and about $\Delta f_\omega$6 average accuracy and roughly $\Delta f_\omega$7 last-task accuracy on CIFAR100 across settings (Guo et al., 25 Mar 2025). The ablation reports that the feature optimization adapter helps but can increase forgetting if used alone, stacking-based parameter fusion substantially improves both average and final accuracy, the dynamic balance factor further boosts performance, and distillation mainly improves retention of old tasks (Guo et al., 25 Mar 2025).

In imbalanced semi-supervised learning, L2AC improves both MixMatch and FixMatch across multiple imbalance ratios and distribution settings. On CIFAR-10 with $\Delta f_\omega$8, it improves FixMatch from 71.5/66.8 to 82.1/81.5 in bACC/GM and MixMatch from 64.8/49.0 to 76.6/75.7 (Wang et al., 2022). Diagnostic evidence includes better pseudo-label recall for minority classes, more balanced per-class recall, and a predicted class distribution closer to the upper bound model trained on balanced fully labeled data (Wang et al., 2022).

For open-vocabulary camouflaged object segmentation, the classifier-centric method improves OVCoser on OVCamo from 0.443 to 0.493 in cIoU, from 0.579 to 0.658 in cSm, and reduces cMAE from 0.336 to 0.239. It also improves classification relative to Alpha-CLIP from 74.60% to 79.75% with GT mask and from 69.88% to 76.85% with all-black mask (Zhang et al., 29 Sep 2025). The oracle-classifier ablation shows that stronger classification is associated with dramatically improved segmentation, indicating that the semantic recognition stage is a bottleneck (Zhang et al., 29 Sep 2025).

In fairness-aware face recognition, GAC reduces biasness on RFW from 1.11 to 0.58 compared with ArcFace while improving average accuracy from 94.64 to 95.21, and remains competitive on LFW, IJB-A, and IJB-C (Gong et al., 2020). In FPGA deployment, RTANN achieves 84.66% on German Credit, 87.06% on Diabetes, and 94.48% on Vehicle Silhouette, outperforming the best individual neural network in each case and reporting resource savings through partial reconfiguration relative to static deployment (Bouazzaoui et al., 2023). CaCT reports smaller prediction set sizes while maintaining coverage $\Delta f_\omega$9 across balanced and long-tailed benchmarks, with representative THR results including set size $$z'_{t-1} = (1-\lambda)\,\phi_{fc}(z_{t-1}) + \lambda z_{t-1}.$$0 and coverage gap $$z'_{t-1} = (1-\lambda)\,\phi_{fc}(z_{t-1}) + \lambda z_{t-1}.$$1 on CIFAR100 and set size $$z'_{t-1} = (1-\lambda)\,\phi_{fc}(z_{t-1}) + \lambda z_{t-1}.$$2 with coverage gap $$z'_{t-1} = (1-\lambda)\,\phi_{fc}(z_{t-1}) + \lambda z_{t-1}.$$3 on ImageNet (Marani et al., 14 Jan 2026).

6. Limitations, misconceptions, and open directions

The literature also makes clear that classifier-centricity is not a universal remedy. Some methods depend on accurate auxiliary information. Group Adaptive Classifier performs best with ground-truth demographics, degrades with estimated demographics, and performs worst with random demographics, which confirms that demographic conditioning must be meaningful (Gong et al., 2020). Domain classifier banks rely on pseudo labels from a mean teacher; the paper therefore introduces teacher-confidence gating and entropy-based weighting precisely because noisy pseudo labels can destabilize class-conditioned alignment (Tang et al., 2020). In imbalanced semi-supervised learning, the paper explicitly notes that without the bi-level separation, the bias attractor can help only weakly (Wang et al., 2022).

Another misconception is that classifier-centric adaptation always minimizes computational cost. Some instances do, such as FPGA dynamic classifier selection and the lightweight text adapter with about $$z'_{t-1} = (1-\lambda)\,\phi_{fc}(z_{t-1}) + \lambda z_{t-1}.$$4M trainable parameters (Bouazzaoui et al., 2023, Zhang et al., 29 Sep 2025). Others shift complexity rather than eliminate it. CaCT introduces class-wise augmented-Lagrangian updates and differentiable approximations for indicators and quantiles during training (Marani et al., 14 Jan 2026). CLIP-based continual learning adds low-rank decomposition, stacking, and MMD/LDA-derived balancing (Guo et al., 25 Mar 2025). ACL, while orthogonal to classifier-centric continual-learning methods, reports that adapting the entire PTM increases GPU memory usage by about 7 GB under its setup (Lu et al., 4 Jun 2025).

A further point of clarification is that “adaptive” does not always mean online parameter modification after deployment. In RTANN, adaptation is sample-specific classifier selection and hardware reconfiguration, not changing network weights online (Bouazzaoui et al., 2023). In strategic classification, adaptation refers to the anticipated best response of decision subjects to the deployed classifier (Chen et al., 2020). In AWDF, adaptation occurs across cascade levels through instance reweighting, and the method is described as more flexible than hard confidence screening because it assigns continuous weights rather than removing instances (Utkin et al., 2019).

Taken together, these works indicate that classifier-centric adaptive framework is best understood as a research program concerned with the evolution, decomposition, routing, or regularization of classifiers under non-stationarity, imbalance, fairness constraints, domain shift, or deployment feedback. A plausible implication is that the framework becomes most valuable when the dominant failure mode is not generic representation quality alone, but the way a decision rule absorbs skew, forgets prior boundaries, miscalibrates uncertainty, or routes responsibility across experts.