Stage-wise Adaptive Label Distribution Learning

• The paper proposes a novel framework that models label ambiguity by partitioning the age range into stages, rather than using a fixed global variance.
• It introduces a two-phase strategy with Stage-wise Adaptive Variance (SAV) and Stage-wise Adaptive Weighted loss (SAW) to finely adjust uncertainty and supervision based on age groups.
• Empirical results on datasets like MORPH-II and FG-NET demonstrate that this method significantly reduces MAE and improves cumulative scores compared to traditional approaches.

Stage-wise Adaptive Label Distribution Learning (SA-LDL) is a label distribution learning framework for facial age estimation that models label ambiguity as varying across age stages rather than as a single global phenomenon. In the broader LDL paradigm, each instance is associated with a distribution of description degrees over labels, with single-label and multi-label learning recovered as special cases (Geng, 2014). SA-LDL instantiates this idea for ages $0$–$100$, arguing that ambiguity is relatively homogeneous within some age ranges but changes across ranges, and therefore should be represented by stage-specific variances and stage-specific loss weights rather than a fixed Gaussian spread over the entire age axis (Wu et al., 30 Aug 2025). Earlier age-estimation LDL in speech already showed that Gaussian label distributions, KL-based fitting, expectation decoding, and uncertainty regularization can outperform pure regression or pure classification, but it used a fixed ambiguity parameter and did not define SA-LDL explicitly (Si et al., 2022).

1. Conceptual basis in label distribution learning

LDL represents an instance by a label distribution

$$D_i = \{ d_{\mathbf{x}_i}^{y_1}, d_{\mathbf{x}_i}^{y_2}, \dots, d_{\mathbf{x}_i}^{y_c} \},$$

with non-negativity and simplex constraints, so that each label contributes to the description of the instance to a different degree (Geng, 2014). In this formulation, single-label learning corresponds to a one-hot distribution and multi-label learning can be expressed as a distribution over the positive label set. The central consequence is that ambiguity is not treated as annotation noise external to the target; it is encoded directly in the target itself.

Facial age estimation is a particularly natural LDL problem because adjacent ages are visually correlated and the degree of ambiguity is not uniform across the age axis. SA-LDL defines “stage-wise” as partitioning the age range $0$–$100$ into contiguous stages and modeling ambiguity per stage rather than globally. The motivating examples are explicit: early childhood is described as comparatively distinguishable, young adulthood such as ages $22$–$28$ as highly ambiguous, and later adulthood as following a different ambiguity pattern (Wu et al., 30 Aug 2025). This departs from fixed-form LDL, which assumes a single variance for all ages, and also from fully per-sample adaptation, which the SA-LDL paper characterizes as potentially noisy and overfitting-prone.

The immediate precursor in age estimation is fixed-variance LDL for speaker age. There, each scalar age label is converted into a Gaussian-smoothed age distribution, the model is trained by KL divergence, and age is decoded by the expectation of the predicted distribution. That framework also combines distribution fitting with $L_1$ age regression and a variance penalty on the predicted distribution, thereby coupling ambiguity modeling and uncertainty minimization, but it retains a single $\sigma$ across ages (Si et al., 2022). SA-LDL can therefore be understood as replacing globally uniform ambiguity with stage-structured ambiguity.

2. Stage-wise ambiguity modeling and the SAV mechanism

SA-LDL is grounded in an embedding-similarity analysis performed with an EfficientNetV2 backbone. For an anchor image at age $a$, the method examines

$100$0

The reported observation is that similarity declines with distance from $100$1, but the shape of the similarity curve changes across age ranges. Within some ranges, the effective ambiguity is relatively constant; across ranges, it shifts. The paper describes these as piecewise-stable behaviors and takes them as evidence that ambiguity is structured by age stage rather than by a single global rule (Wu et al., 30 Aug 2025).

On this basis, SA-LDL divides the age range $100$2 into $100$3 stages. These are not simple 10-year bins. Instead, ages are grouped by K-means clustering over labels from MORPH-II, FG-NET, UTK, and CACD after feature extraction with EfficientNetV2, so that each stage is intended to capture similar facial feature distribution patterns. For each stage $100$4, SA-LDL maintains a stage-specific standard deviation $100$5.

For an image $100$6 with chronological age $100$7, the ground-truth label distribution becomes

$100$8

The network outputs logits $100$9, converted by softmax into

$$D_i = \{ d_{\mathbf{x}_i}^{y_1}, d_{\mathbf{x}_i}^{y_2}, \dots, d_{\mathbf{x}_i}^{y_c} \},$$0

Distribution fitting is performed with

$$D_i = \{ d_{\mathbf{x}_i}^{y_1}, d_{\mathbf{x}_i}^{y_2}, \dots, d_{\mathbf{x}_i}^{y_c} \},$$1

The Stage-wise Adaptive Variance mechanism, SAV, learns the stage-specific variances $$D_i = \{ d_{\mathbf{x}_i}^{y_1}, d_{\mathbf{x}_i}^{y_2}, \dots, d_{\mathbf{x}_i}^{y_c} \},$$2 through an outer-loop validation procedure. The algorithm maintains model parameters $$D_i = \{ d_{\mathbf{x}_i}^{y_1}, d_{\mathbf{x}_i}^{y_2}, \dots, d_{\mathbf{x}_i}^{y_c} \},$$3, stage variances $$D_i = \{ d_{\mathbf{x}_i}^{y_1}, d_{\mathbf{x}_i}^{y_2}, \dots, d_{\mathbf{x}_i}^{y_c} \},$$4, and the best validation $$D_i = \{ d_{\mathbf{x}_i}^{y_1}, d_{\mathbf{x}_i}^{y_2}, \dots, d_{\mathbf{x}_i}^{y_c} \},$$5 error. At each epoch, the model is trained with $$D_i = \{ d_{\mathbf{x}_i}^{y_1}, d_{\mathbf{x}_i}^{y_2}, \dots, d_{\mathbf{x}_i}^{y_c} \},$$6 using the stage-appropriate $$D_i = \{ d_{\mathbf{x}_i}^{y_1}, d_{\mathbf{x}_i}^{y_2}, \dots, d_{\mathbf{x}_i}^{y_c} \},$$7 for each sample, then evaluated on a validation set; when the validation $$D_i = \{ d_{\mathbf{x}_i}^{y_1}, d_{\mathbf{x}_i}^{y_2}, \dots, d_{\mathbf{x}_i}^{y_c} \},$$8 improves, the current $$D_i = \{ d_{\mathbf{x}_i}^{y_1}, d_{\mathbf{x}_i}^{y_2}, \dots, d_{\mathbf{x}_i}^{y_c} \},$$9 and $0$0 are retained as the best parameters. This makes variance selection a stage-level search guided by age estimation accuracy rather than by a fixed prior (Wu et al., 30 Aug 2025).

A compact expression for the resulting stage-conditioned target is

$0$1

where $0$2 denotes the stage containing age $0$3. The substantive change relative to earlier LDL is that the spread of label ambiguity is no longer global; it is indexed by stage.

3. Stage-wise Adaptive Weighted loss and task coupling

SA-LDL does not treat facial age estimation as only a distribution-matching problem. It explicitly combines label distribution learning, age classification, and age regression, then modulates the balance between them by stage. This component is called Stage-wise Adaptive Weighted loss, or SAW (Wu et al., 30 Aug 2025).

The component losses are: $0$4 for classification,

$0$5

for regression, and

$0$6

for label-distribution fitting. The final objective is

$0$7

Here $0$8 is stage-wise adaptive. In the paper’s interpretation, stages with stronger ambiguity receive higher $0$9, shifting emphasis toward $100$0; stages with weaker ambiguity receive lower $100$1, shifting emphasis toward $100$2. The $100$3 term remains at a small constant weight and acts as a regularizer linking the distributional prediction to the scalar age (Wu et al., 30 Aug 2025). The expected age is computed from the predicted distribution as $100$4, and this quantity supplies the regression target.

This design places SA-LDL midway between two earlier lines. Fixed-form LDL used KL divergence alone or a fixed Gaussian variance; speaker-age LDL already combined KL, regression on the expectation, and a variance-related uncertainty term, showing that age estimation benefits from hybridization of classification-like and regression-like objectives (Si et al., 2022). SA-LDL extends that hybridization by making the balance itself stage-dependent rather than globally fixed.

4. Architecture and optimization pipeline

SA-LDL uses EfficientNetV2 as its backbone network. Facial images are resized to $100$5, the backbone produces feature embeddings $100$6, and a final classification head outputs logits for age categories $100$7–$100$8. Softmax then yields the predicted label distribution $100$9, from which the expected age can be decoded (Wu et al., 30 Aug 2025).

Embeddings serve two distinct roles. First, they are used in the pre-training analysis of cosine similarity across ages, which motivates stage-wise ambiguity modeling. Second, they are used in K-means clustering to construct the $22$0 age stages. This gives the stage partition an empirical basis in feature-space behavior rather than imposing a purely chronological segmentation.

Optimization is performed with SGD, a base learning rate of $22$1, and batch size $22$2. The implementation is in PyTorch and was run on an NVIDIA RTX 3090 GPU. Training proceeds in two linked phases. SAV first learns stage-specific variances $22$3 using validation $22$4 performance. Those learned variances are then used in downstream training with the SAW objective, where the stage-specific $22$5 controls the relative contribution of $22$6 and $22$7 while $22$8 remains constant (Wu et al., 30 Aug 2025).

The architecture is therefore simple in topology but structured in supervision. The adaptation resides not in multiple network heads or a cascade, but in the label distribution assigned to each age stage and in the stage-conditioned weighting of the coupled loss terms.

5. Empirical results and ablation evidence

The principal experiments are reported on MORPH-II, FG-NET, UTK, and CACD. MORPH-II is described as age $22$9–$28$0 under controlled conditions; FG-NET as age $28$1–$28$2 and small, in-the-wild; UTK as $28$3–$28$4 and unconstrained; CACD as cross-age celebrity (Wu et al., 30 Aug 2025).

The paper reports competitive performance across these datasets. The abstract states that SA-LDL achieves “MAE of 1.74 and 2.15 on the MORPH-II and FG-NET datasets.” In the experimental summary, the reported numbers are SA-LDL: MAE $28$5, CS $28$6 on MORPH-II; MAE $28$7, CS $28$8 on FG-NET; MAE $28$9 on UTK; and MAE $L_1$0 on CACD (Wu et al., 30 Aug 2025).

Ablation results isolate the two defining components. For stage-wise variance alone, replacing a predefined $L_1$1 with SAV improves MORPH-II from $L_1$2 to $L_1$3 in MAE / CS, FG-NET from $L_1$4 to $L_1$5, UTK from $L_1$6 to $L_1$7, and CACD from $L_1$8 to $L_1$9, with the paper noting a dramatic CS improvement on CACD despite an apparent typo in the baseline entry. For the weighted-loss component, replacing pure $\sigma$0, pure $\sigma$1, or a simple $\sigma$2 combination with SAW yields further MAE reductions: on MORPH-II from $\sigma$3 to $\sigma$4, on FG-NET from $\sigma$5 to $\sigma$6, on UTK from $\sigma$7 to $\sigma$8, and on CACD from $\sigma$9 to $a$0. The full configuration, combining SAV and SAW, reports MAE $a$1 on MORPH-II, $a$2 on FG-NET, $a$3 on UTK, and $a$4 on CACD, outperforming the baseline, SAV-only, and SAW-only variants (Wu et al., 30 Aug 2025).

These results matter because they separate two claims. First, stage-wise ambiguity modeling alone improves age estimation. Second, stage-wise coupling of distributional and categorical supervision adds further gains. The paper’s interpretation is that the two mechanisms are complementary: SAV aligns label-distribution spread with ambiguity patterns, whereas SAW aligns the optimization objective with stage-specific ambiguity.

6. Position within the LDL literature

SA-LDL sits within a larger LDL literature that spans foundational formulations, age-estimation-specific models, learning from logical labels, multi-view label enhancement, uncertainty-aware latent spaces, and online adaptive supervision. The original LDL formulation defines description degrees over a complete label set and proposes specialized algorithms that optimize distributional objectives directly, emphasizing KL divergence and multi-measure evaluation (Geng, 2014). SA-LDL inherits this general view but imposes a stage structure over the label axis.

Earlier age-estimation LDL in speech provides the clearest technical antecedent. There, each age label is converted into a Gaussian distribution over adjacent ages, the model is trained by KL divergence, age is decoded by the expectation of the predicted distribution, and hybrid losses combine distribution fitting, $a$5 regression, and uncertainty minimization (Si et al., 2022). What SA-LDL adds is not the basic LDL machinery but the claim that ambiguity should be modeled by stage-specific $a$6 and stage-specific $a$7.

Related LDL work from logical labels addresses a different bottleneck: the absence of full label distributions. Direct LDL from logical labels unifies label enhancement and LDL in a joint model, constrains inferred distributions so that invalid labels receive zero mass, and couples recovered distributions to predictions by KL divergence (Jia et al., 2023). Contrastive Label Enhancement instead aligns features and logical labels in a shared projection space and adds explicit label-attribute consistency constraints before recovering distributions (Wang et al., 2023). These methods concern how to obtain or refine label distributions when full distributions are unavailable, whereas SA-LDL assumes age labels and focuses on how the ambiguity structure over those labels should vary across stages.

Another adjacent line models uncertainty in the label space itself. The Latent Label Distribution Grid constructs a label-correlation tensor from pairwise label differences, expands each difference into a Gaussian vector, and applies Tucker-based low-rank reconstruction to produce a low-noise representation before predicting the final distribution (Sun et al., 27 May 2025). This suggests a different route to adaptive LDL: uncertainty can be captured either by stage-conditioned Gaussian targets, as in SA-LDL, or by a latent uncertainty-aware grid. A plausible implication is that stage-wise ambiguity modeling and latent uncertainty modeling are complementary rather than mutually exclusive.

Dynamic LDL during training has also appeared in image classification. Online LDL treats soft labels as evolving supervision, generated for augmented inputs and shown to improve both calibration and generalization relative to hard labels or knowledge distillation in several CNN settings (Hong et al., 2023). Although that work does not define SA-LDL, it reinforces the broader principle that label distributions need not be static. SA-LDL specializes that principle to age estimation by making adaptation depend on age stage.

7. Limitations, common misunderstandings, and extensions

A recurrent misunderstanding is that ambiguity in age estimation must be handled either by one global Gaussian variance or by fully per-sample adaptation. SA-LDL is motivated precisely as an intermediate position. The paper argues that fixed variance overlooks the varying degrees of ambiguity across the age spectrum, whereas per-age or per-sample adaptation can be noisy, unstable, and overfitting-prone. Its central claim is that ambiguity is approximately constant within stages and different between stages (Wu et al., 30 Aug 2025).

The method nevertheless has explicit limitations. It depends on stage partitioning produced by K-means clustering over multiple datasets; if the partition is suboptimal or biased by dataset composition, the learned $a$8 and $a$9 may not reflect the true ambiguity structure. It assumes that $100$00 stages are sufficient. It may inherit dataset bias, because ambiguity patterns can depend on ethnicity, camera conditions, health, or lifestyle. SAV introduces an outer-loop search over stage variances, and SAW requires stage-wise weights $100$01, so the training procedure is more complex than fixed-variance LDL. The paper also notes that it does not specify a fully explicit learning rule for $100$02, which leaves part of the stage-wise weighting mechanism implicit (Wu et al., 30 Aug 2025).

The extensions proposed in the paper are broader than facial age estimation. They include other ordered and ambiguous prediction tasks such as age-group classification, time-to-event estimation, and progression stages in medical imaging; emotion intensity estimation; apparent age and other subjective human-perception ratings; dynamic or learned stage boundaries via differentiable clustering or gating; and hierarchical or multi-scale ambiguity modeling (Wu et al., 30 Aug 2025). These proposals remain speculative within the source, but they are consistent with a general LDL trend toward progressively richer ambiguity models, whether through joint logical-label learning (Jia et al., 2023), multi-view enhancement (Wang et al., 2023), latent uncertainty grids (Sun et al., 27 May 2025), or online evolving label distributions (Hong et al., 2023).

In that sense, SA-LDL is best understood not as a generic synonym for adaptive LDL, but as a specific hypothesis about the geometry of ambiguity in age estimation: ambiguity is structured by age stage, and both the target distribution spread and the loss balance should follow that structure.