---
title: Ambiguity-guided Learnable Distribution Calibration
url: https://www.emergentmind.com/topics/ambiguity-guided-learnable-distribution-calibration-aldc
type: topic
---

# Ambiguity-guided Learnable Distribution Calibration

Ambiguity-guided Learnable Distribution Calibration (ALDC) is a strategy introduced for Generalized Semi-Few-Shot Class-Incremental Learning (GSemi-FSCIL) that uses ambiguity in unlabeled data as a routing signal for pseudo-label selection and distribution correction. In the formulation of ALDC, unlabeled samples may originate from base classes, current novel classes, or previously introduced novel classes, and the method addresses the resulting ambiguity by combining a dynamic thresholding mechanism with base-to-novel distribution calibration and synthetic feature generation [2507.23237]. A conceptually related but terminologically distinct line appears in natural language understanding, where ambiguity-aware self-distillation and confidence re-calibration are used to learn softer label distributions; that work does not use the term “ALDC” explicitly [2406.09719].

## 1. Problem setting and formal scope

In standard Few-Shot Class-Incremental Learning (FSCIL), training begins with a base session containing many labeled examples, followed by incremental sessions in which each new class has only a few labeled samples. Semi-supervised Few-Shot Class-Incremental Learning (Semi-FSCIL) augments this setting with unlabeled data, but existing methods commonly assume that the unlabeled set contains only samples from the current novel classes. ALDC is defined in response to a broader and explicitly more realistic formulation, Generalized Semi-FSCIL (GSemi-FSCIL), in which the unlabeled set at session $t$ is drawn from the full seen label space [2507.23237]:

$$
C_t^u = \{C_0 \cup C_1 \cup \cdots \cup C_t\}.
$$

This redefinition changes the role of unlabeled data. Under GSemi-FSCIL, an unlabeled sample may belong to a base class, the current novel class, or an earlier novel class. The paper identifies the resulting difficulty as class ambiguity in unlabeled data. Because base classes have many samples while novel classes are few-shot, the learned classifier and feature distribution become biased toward base classes. The reported consequences are that some unlabeled samples from novel classes are misclassified as base classes, some unlabeled base samples are confused with novel classes, pseudo-labeling can reinforce wrong decisions, and few-shot novel class prototypes become biased by sparse support [2507.23237].

Within this setting, high-ambiguity samples are defined as unlabeled samples that are similarly close to both base and novel class weights. ALDC is designed specifically for this case, rather than treating ambiguity as noise to be discarded.

## 2. Architectural decomposition of ALDC

The ALDC pipeline has three main components: base pre-training, Ambiguity-guided Learnable Threshold (ALT), and Base-to-Novel (B2N) Distribution Calibration [2507.23237]. The structure is organized around the distinction between easy unlabeled samples, which can be pseudo-labeled directly, and hard unlabeled samples, which are routed into a calibration pathway.

| Component | Function | Output |
|---|---|---|
| Base pre-training | Train on labeled base dataset $D_0$ and store base-class statistics | Base means and covariances |
| ALT | Separate high-confidence from high-ambiguity unlabeled samples | Pseudo-label or calibration route |
| B2N calibration | Refine few-shot novel distributions using base-class statistics | Calibrated Gaussians and synthetic features |

At session $0$, the feature extractor and classifier are trained on the labeled base dataset $D_0$. The resulting model stores base-class feature statistics that are later reused during incremental learning. In later sessions, ALT evaluates each unlabeled sample by comparing its similarity to base and novel classifier weights, and the outcome determines whether the sample is treated as high-confidence or high-ambiguity. High-confidence samples receive pseudo-labels directly; high-ambiguity samples are handled by B2N calibration, which uses base-class statistics to correct biased feature distributions for few-shot novel classes [2507.23237].

The central design choice is not merely to screen unreliable pseudo-labels, but to transform uncertainty into an explicit signal for distribution estimation.

## 3. Ambiguity-guided Learnable Threshold

ALT computes, for each unlabeled sample $x_i$, a feature representation $f_\theta(x_i)$ and compares it with both base and novel classifier weights via cosine similarity [2507.23237]:

$$
S_i^{b,n} = \max\big(\cos(f_\theta(x_i), w^{b,n})\big).
$$

The method then measures the absolute gap between the maximum base confidence and the maximum novel confidence. If the gap is small, the sample is treated as ambiguous; if it is large, the sample is easier to assign. The threshold is learnable and dynamic rather than fixed:

$$
\tau = \frac{1}{N_u}\sum_{i=1}^{N_u} \left|S_i^b - S_i^n\right| + m,
$$

where $N_u$ is the number of unlabeled samples and $m$ is a smoothing coefficient [2507.23237].

The decision rule is dual-branch. If $\left|S^b - S^n\right| > \tau$, the sample is considered high-confidence and receives a pseudo-label from the class with maximum similarity. Otherwise, it is treated as high-ambiguity and passed to B2N calibration. The paper characterizes this as a dynamic mechanism that adapts to changing uncertainty across sessions, in contrast to a static threshold. In the ablation studies, dropping ambiguous samples harms performance, static thresholding is better than dropping, and dynamic ALT performs best [2507.23237].

A common misunderstanding is to regard ALT as a confidence-only filter. The method is more specific: it uses the base-versus-novel confidence gap as an ambiguity signal and routes uncertain samples into a separate calibration procedure rather than discarding them.

## 4. Base-to-Novel distribution calibration

B2N calibration addresses the instability of few-shot novel-class statistics by transferring structure from abundant base classes to sparse novel classes. The paper first computes class-wise feature means and covariances as empirical Gaussian statistics [2507.23237]:

$$
\begin{aligned}
\mu_c &= \frac{1}{N_c^t}\sum_{i=1}^{N_c^t} f_\theta(x_i), \\
\Sigma_c &= \frac{1}{N_c^t}\sum_{x_j\in D_t} \big(f_\theta(x_j)-\mu_c\big)\big(f_\theta(x_j)-\mu_c\big)^T.
\end{aligned}
$$

For ambiguous unlabeled samples, the method assumes latent correlation between a base class and a novel class. Because few-shot novel classes have unreliable statistics, the novel distribution is calibrated using base-class statistics. The calibrated statistics are [2507.23237]:

$$
{\mu}_{c}^{\prime}=\frac{\sum_{i\in D_0}\mu_i+\mu_c}{N_c^t+1}, \qquad
{\Sigma}_{c}^{\prime}=\frac{\sum_{i\in D_0}\Sigma_i+\Sigma_c}{N_c^t+1}+\alpha,
$$

where $\alpha$ is a hyperparameter controlling covariance smoothing.

The text states that the base-class distribution is selected or transferred to be more similar to the novel class according to cosine similarity, so that the novel representation is anchored by richer base knowledge. After calibration, synthetic features are sampled from the Gaussian

$$
x \sim \mathcal{N}(\mu_c', \Sigma_c').
$$

These generated features are labeled and added to the training set, effectively turning an $N$-way $K$-shot problem into an $N$-way $(K+10)$-shot setting in the implementation example reported in the paper [2507.23237].

Functionally, B2N targets biased novel feature distributions. Few-shot novel classes have unstable and under-estimated statistics, and their class prototypes can be unreliable because of too few labeled samples. B2N reduces this bias by injecting richer base-class statistical structure into novel distribution estimation.

## 5. Session-wise training and inference

The training procedure is divided into a base session and incremental sessions. In the base session, the backbone and classifier are trained on $D_0$, and base class means and covariances are stored. In each incremental session $n>0$, novel class weights are initialized using labeled few-shot data $D_n^l$; each unlabeled sample is evaluated through base and novel similarities; the similarity gap is compared with the current threshold; high-confidence samples are pseudo-labeled; high-ambiguity samples undergo B2N calibration; the threshold is updated dynamically; novel means and covariances are updated; synthetic features are sampled from calibrated Gaussians; and the classifier is retrained or updated using the union of labeled, pseudo-labeled, and generated samples [2507.23237].

Inference remains standard classification over all seen classes so far. The novelty lies in the construction of the training set and in the way uncertainty is used. ALDC does not simply apply pseudo-labeling to all unlabeled data, nor does it simply reject uncertain samples. Instead, it separates easy and difficult unlabeled samples and assigns them different learning roles.

This design addresses two coupled problems described in the paper. The first is ambiguous unlabeled samples: naive pseudo-labeling can be harmful because unlabeled data may come from base or novel classes. The second is biased novel feature distributions: few-shot novel classes have sparse support, and their learned distributions can be poorly estimated. ALT acts on the former, and B2N acts on the latter, with the combined method treating ambiguity as a source of calibration information rather than only as a failure mode [2507.23237].

## 6. Empirical results, baselines, and ablations

Experiments are reported on mini-ImageNet, CIFAR-100, and CUB-200. The comparison set includes FSCIL and Semi-FSCIL baselines such as iCaRL, TOPIC, CEC, FACT, LIMIT, MCNet, SS-iCaRL, SS-NCM-CNN, Semi-CEC, Semi-Limit, Us-KD, and UaD-CE. Under GSemi-FSCIL comparisons, Semi-FSCIL baselines are marked with $^*$ to indicate evaluation under the generalized unlabeled setting. The primary metrics are classification accuracy per session and average accuracy across sessions [2507.23237].

On mini-ImageNet under GSemi-FSCIL, ALDC achieves an average accuracy of $69.77\%$, compared with the best previous Semi-FSCIL baseline UaD-CE$^*$ at $58.26\%$, for an improvement of $+11.51\%$. The paper also reports that ALDC surpasses prior Semi-FSCIL methods in the standard Semi-FSCIL setting, with a reported improvement of $+10.93\%$ over the best prior method. Its session-wise performance on mini-ImageNet under GSemi-FSCIL is
$$
82.45,\ 78.85,\ 76.75,\ 72.87,\ 69.98,\ 64.54,\ 64.66,\ 60.19,\ 58.17.
$$

| Result | Value | Context |
|---|---:|---|
| Average accuracy | 69.77% | ALDC on mini-ImageNet under GSemi-FSCIL |
| Best previous Semi-FSCIL baseline | 58.26% | UaD-CE$^*$ |
| Improvement | +11.51% | GSemi-FSCIL comparison |
| Baseline all-class accuracy | 66.27 | ALT/B2N ablation |
| ALT only | 67.82 | ALT/B2N ablation |
| B2N only | 67.87 | ALT/B2N ablation |
| ALT + B2N | 69.77 | ALT/B2N ablation |

The ablation studies attribute complementary roles to the two main modules. ALT improves pseudo-label selection and protects base-class performance, while B2N improves novel-class generalization by mitigating overfitting and biased statistics. The reported all-class accuracies on mini-ImageNet are $66.27$ for the baseline, $67.82$ for ALT only, $67.87$ for B2N only, and $69.77$ for ALT + B2N [2507.23237].

The study of unlabeled sample count evaluates $N_u \in \{25, 50, 75, 125\}$. The best reported performance in the ablation table is at $125$ unlabeled samples, with average accuracy $70.80$, although $50$ unlabeled samples are used for fair comparative experiments. Hyperparameter sensitivity is analyzed for $m$, the smoothing coefficient in the threshold, and $\alpha$, the covariance smoothing parameter in B2N. The paper reports that larger $m$ smooths the threshold and helps stability, and that a proper $\alpha$ improves novel-class performance. The t-SNE visualization shows cleaner class clusters, sharper decision boundaries, and better-separated novel and base classes compared with confidence-only pseudo-labeling. The paper also presents high-ambiguity examples of unlabeled samples that are semantically similar to both base and novel classes [2507.23237].

## 7. Relation to ambiguity-aware distribution calibration in other domains

Although ALDC is explicitly named in the GSemi-FSCIL context, a closely aligned ambiguity-guided calibration logic appears in natural language understanding in “Self-Knowledge Distillation for Learning Ambiguity” [2406.09719]. That work studies NLU classification problems in which a single input can legitimately map to multiple labels, and it addresses the tendency of pretrained language models fine-tuned with one-hot labels to become over-confident on ambiguous samples.

The method in that paper is a self-knowledge distillation framework based on the empirical finding that lower layers preserve sample-label relationship information better than higher layers. A warm-up phase attaches classifiers to all layers and identifies a source layer whose validation entropy drops sharply after some early fine-tuning epoch; the model is then reset and trained in two coupled steps, with the main classifier learning from the source layer’s softer distribution and the source layer updated using the main classifier. After this ambiguity-distribution learning phase, the method sorts samples by an ambiguity score computed from the confidence of the ground-truth label across layers from the source layer to the top layer, selects the bottom $10\%$ as “extremely ambiguous,” and performs a one-epoch re-calibration step on that subset [2406.09719].

The re-calibration loss is

$$
L_{re\text{-}calibration} = 0.5*L_{ce}(\bar{y}_{main}, y) + 0.5*L_{ce}(\bar{y}_{main}, u),
$$

where $u$ is the uniform distribution. The stated purpose is to reduce unnecessarily strengthened confidence for ambiguous samples without driving the model toward near-uniform predictions on ordinary unseen samples. On ChaosSNLI and ChaosMNLI, the reported method variants LAD and LAD + RC improve distribution quality, especially KL divergence; Table 3 in the paper reports average-difference improvements for mispredicted samples from $0.5110$ to $0.4655$ on ChaosSNLI and from $0.4479$ to $0.4122$ on ChaosMNLI when RC is added [2406.09719].

The relationship between the two papers is conceptual rather than terminological. The 2024 NLU work does not explicitly mention “Ambiguity-guided Learnable Distribution Calibration,” but it shares several structural ideas: ambiguity-aware sample selection, distribution learning rather than hard-label optimization alone, and targeted confidence re-calibration rather than uniform global smoothing. A plausible implication is that ALDC belongs to a broader family of methods in which ambiguity is treated as an informative property of data distributions rather than only as uncertainty to suppress.

Source: https://www.emergentmind.com/topics/ambiguity-guided-learnable-distribution-calibration-aldc