Generalized Pseudo-Label Robust Loss

• GPR Loss is a robust loss formulation for single positive multi-label learning that distinguishes between confirmed positives, unlabeled data, and pseudo-labels.
• It combines case-specific loss branches, adaptive weighting functions, and a cardinality regularizer to effectively mitigate pseudo-label noise.
• Empirical results show that integrating GPR Loss consistently improves mean Average Precision across diverse pseudo-label sources in the AEVLP framework.

Generalized Pseudo-Label Robust Loss (GPR Loss) is a loss formulation for Single Positive Multi-Label Learning (SPML) that extends prior robust SPML objectives to incorporate explicit pseudo-labels produced by an external method. It was introduced in the context of the Adaptive and Efficient Vision-Language Pseudo-Labeling (AEVLP) framework, where each image carries one observed positive label under the Assume Negative convention and may additionally receive pseudo-positive, pseudo-negative, or undefined pseudo-label assignments. GPR Loss is defined over these heterogeneous supervision states by combining case-specific losses, case-specific weights, and a cardinality regularizer, with the stated aim of learning from diverse pseudo-labels while mitigating pseudo-label noise (Tran et al., 28 Aug 2025).

1. Problem setting in SPML

In standard multi-label learning, the training set is

$\mathcal{D} = \{(x_n, y_n)\}_{n=1}^{N},$

where $x_n \in \mathcal{X}$ and $y_n \in \{0,1\}^C$ is a binary label vector over $C$ classes. The model predicts

$p_n = f(x_n)=\sigma(s_n),$

with logits $s_n$ and probabilities $p_{n,i}\in(0,1)$. Under full supervision, the binary cross-entropy objective is

$$\mathcal{L} = -\frac{1}{NC} \sum_{n=1}^N \sum_{i=1}^C y_{n,i}\log p_{n,i} + (1-y_{n,i})\log(1 - p_{n,i}).$$

SPML is an extreme partial-annotation regime in which each image is provided with only one confirmed positive label. The training set is converted into a pseudo multi-label dataset

$$\hat{\mathcal{D}} = \{(x_n,\hat{y}_n)\}_{n=1}^N,\qquad \hat{y}_n \in \{0,1\}^C,\qquad \sum_{i=1}^C \hat{y}_{n,i}=1.$$

Here, $\hat{y}_{n,i}=1$ denotes the observed positive, while the remaining zeros are unknown or unannotated rather than verified negatives. In this regime, ordinary BCE creates false-negative supervision because it interprets almost all zeros as true negatives (Tran et al., 28 Aug 2025).

To recover missing supervision, GPR Loss assumes that each instance is paired with an auxiliary pseudo-label vector

$x_n \in \mathcal{X}$0

where $x_n \in \mathcal{X}$1 is produced by an arbitrary pseudo-labeling method $x_n \in \mathcal{X}$2. For each class $x_n \in \mathcal{X}$3, $x_n \in \mathcal{X}$4 denotes a positive pseudo-label, $x_n \in \mathcal{X}$5 a negative pseudo-label, and $x_n \in \mathcal{X}$6 an undefined entry. This produces four supervision cases for each image-class pair: observed positive, unlabeled without pseudo-label, pseudo-negative, and pseudo-positive. GPR Loss is designed precisely for this four-way decomposition (Tran et al., 28 Aug 2025).

2. Formal definition

GPR Loss generalizes the earlier Generalized Robust Loss (GR Loss) by splitting the Assume Negative case according to the pseudo-label state. Its full form is

$x_n \in \mathcal{X}$7

where $x_n \in \mathcal{X}$8 controls weighting functions, $x_n \in \mathcal{X}$9 is the regularization coefficient, and $y_n \in \{0,1\}^C$0 is a positive-count regularizer.

The case-conditioned loss term is

$y_n \in \{0,1\}^C$1

and the corresponding weight is

$y_n \in \{0,1\}^C$2

The inherited GR components are

$y_n \in \{0,1\}^C$3

$y_n \in \{0,1\}^C$4

with $y_n \in \{0,1\}^C$5 and

$y_n \in \{0,1\}^C$6

The new GPR-specific branches are

$y_n \in \{0,1\}^C$7

$y_n \in \{0,1\}^C$8

together with

$y_n \in \{0,1\}^C$9

and

$C$0

The regularizer is

$C$1

where $C$2 is the expected number of positive labels per image and

$C$3

The paper writes the expected positive count as

$C$4

The functions $C$5 and $C$6 are implemented with stopped gradients with respect to $C$7 (Tran et al., 28 Aug 2025).

3. Loss decomposition and robustness mechanisms

The structure of GPR Loss is explicitly asymmetric across supervision sources. Confirmed positives are trained with

$C$8

so observed positives are fully trusted. Unresolved unlabeled entries, defined by $C$9 and $p_n = f(x_n)=\sigma(s_n),$0, retain the GR Loss branch $p_n = f(x_n)=\sigma(s_n),$1, which mixes positive-like and negative-like penalties through $p_n = f(x_n)=\sigma(s_n),$2. This preserves the SPML-specific treatment of false negatives under Assume Negative (Tran et al., 28 Aug 2025).

Pseudo-negatives use standard negative BCE,

$p_n = f(x_n)=\sigma(s_n),$3

but only after multiplication by

$p_n = f(x_n)=\sigma(s_n),$4

Pseudo-positives are not treated as hard positives. Instead, GPR uses

$p_n = f(x_n)=\sigma(s_n),$5

which is exactly BCE with soft target $p_n = f(x_n)=\sigma(s_n),$6. The corresponding weight

$p_n = f(x_n)=\sigma(s_n),$7

adds both adaptive reweighting and clipping. The paper presents this as the clearest robustness mechanism for noisy pseudo-positives, because it avoids treating them as hard truth while preventing either vanishing or dominating influence (Tran et al., 28 Aug 2025).

The regularizer

$p_n = f(x_n)=\sigma(s_n),$8

constrains the model’s average positive mass and is intended to stabilize training when pseudo-labels increase the tendency to overpredict positives. This couples local pseudo-label handling with a global cardinality prior.

Several special cases are explicit. If no pseudo-labels are provided, meaning $p_n = f(x_n)=\sigma(s_n),$9 for all unknown entries, then GPR collapses to GR Loss up to the regularization behavior. If $s_n$0, then

$s_n$1

which becomes hard-positive BCE; if $s_n$2, then

$s_n$3

which acts like a negative target. The paper also states a theorem that

$s_n$4

tends to

$s_n$5

when

$s_n$6

tends to zero, where $s_n$7 measures pseudo-label confidence on observed positives and $s_n$8 is a validation-set estimate of average label cardinality (Tran et al., 28 Aug 2025).

4. Integration within AEVLP and DAMP

GPR Loss is the loss component of AEVLP, while the pseudo-label generator is DAMP, the Dynamic Augmented Multi-focus Pseudo-labeling procedure. DAMP uses CLIP with both global and local augmented image views. Given image $s_n$9 and class prompt $p_{n,i}\in(0,1)$0, CLIP produces visual embedding $p_{n,i}\in(0,1)$1, text embedding $p_{n,i}\in(0,1)$2, similarity

$p_{n,i}\in(0,1)$3

and class probabilities

$p_{n,i}\in(0,1)$4

A frozen graph $p_{n,i}\in(0,1)$5 is applied on text embeddings through

$p_{n,i}\in(0,1)$6

with graph-convolution updates

$p_{n,i}\in(0,1)$7

For each image $p_{n,i}\in(0,1)$8, DAMP creates one global view

$p_{n,i}\in(0,1)$9

and $$\mathcal{L} = -\frac{1}{NC} \sum_{n=1}^N \sum_{i=1}^C y_{n,i}\log p_{n,i} + (1-y_{n,i})\log(1 - p_{n,i}).$$0 local patch views

$$\mathcal{L} = -\frac{1}{NC} \sum_{n=1}^N \sum_{i=1}^C y_{n,i}\log p_{n,i} + (1-y_{n,i})\log(1 - p_{n,i}).$$1

Using the known single positive $$\mathcal{L} = -\frac{1}{NC} \sum_{n=1}^N \sum_{i=1}^C y_{n,i}\log p_{n,i} + (1-y_{n,i})\log(1 - p_{n,i}).$$2, it defines the local threshold

$$\mathcal{L} = -\frac{1}{NC} \sum_{n=1}^N \sum_{i=1}^C y_{n,i}\log p_{n,i} + (1-y_{n,i})\log(1 - p_{n,i}).$$3

For each class $$\mathcal{L} = -\frac{1}{NC} \sum_{n=1}^N \sum_{i=1}^C y_{n,i}\log p_{n,i} + (1-y_{n,i})\log(1 - p_{n,i}).$$4,

$$\mathcal{L} = -\frac{1}{NC} \sum_{n=1}^N \sum_{i=1}^C y_{n,i}\log p_{n,i} + (1-y_{n,i})\log(1 - p_{n,i}).$$5

and the aggregated local score is

$$\mathcal{L} = -\frac{1}{NC} \sum_{n=1}^N \sum_{i=1}^C y_{n,i}\log p_{n,i} + (1-y_{n,i})\log(1 - p_{n,i}).$$6

The final positive-pseudo-label score is

$$\mathcal{L} = -\frac{1}{NC} \sum_{n=1}^N \sum_{i=1}^C y_{n,i}\log p_{n,i} + (1-y_{n,i})\log(1 - p_{n,i}).$$7

followed by top-$$\mathcal{L} = -\frac{1}{NC} \sum_{n=1}^N \sum_{i=1}^C y_{n,i}\log p_{n,i} + (1-y_{n,i})\log(1 - p_{n,i}).$$8 and threshold-based assignment

$$\mathcal{L} = -\frac{1}{NC} \sum_{n=1}^N \sum_{i=1}^C y_{n,i}\log p_{n,i} + (1-y_{n,i})\log(1 - p_{n,i}).$$9

Negative pseudo-labels are mined from

$$\hat{\mathcal{D}} = \{(x_n,\hat{y}_n)\}_{n=1}^N,\qquad \hat{y}_n \in \{0,1\}^C,\qquad \sum_{i=1}^C \hat{y}_{n,i}=1.$$0

with

$$\hat{\mathcal{D}} = \{(x_n,\hat{y}_n)\}_{n=1}^N,\qquad \hat{y}_n \in \{0,1\}^C,\qquad \sum_{i=1}^C \hat{y}_{n,i}=1.$$1

This produces the ternary pseudo-label vector $$\hat{\mathcal{D}} = \{(x_n,\hat{y}_n)\}_{n=1}^N,\qquad \hat{y}_n \in \{0,1\}^C,\qquad \sum_{i=1}^C \hat{y}_{n,i}=1.$$2 consumed by GPR Loss (Tran et al., 28 Aug 2025).

The training pipeline is epoch-wise. At each epoch, DAMP regenerates pseudo-labels using global and local augmented views, the classifier computes logits $$\hat{\mathcal{D}} = \{(x_n,\hat{y}_n)\}_{n=1}^N,\qquad \hat{y}_n \in \{0,1\}^C,\qquad \sum_{i=1}^C \hat{y}_{n,i}=1.$$3 and probabilities $$\hat{\mathcal{D}} = \{(x_n,\hat{y}_n)\}_{n=1}^N,\qquad \hat{y}_n \in \{0,1\}^C,\qquad \sum_{i=1}^C \hat{y}_{n,i}=1.$$4, and GPR Loss is applied to $$\hat{\mathcal{D}} = \{(x_n,\hat{y}_n)\}_{n=1}^N,\qquad \hat{y}_n \in \{0,1\}^C,\qquad \sum_{i=1}^C \hat{y}_{n,i}=1.$$5. In the main setup, the classifier is a ResNet-50 pretrained on ImageNet, input size is $$\hat{\mathcal{D}} = \{(x_n,\hat{y}_n)\}_{n=1}^N,\qquad \hat{y}_n \in \{0,1\}^C,\qquad \sum_{i=1}^C \hat{y}_{n,i}=1.$$6, the optimizer is Adam, the learning rate is $$\hat{\mathcal{D}} = \{(x_n,\hat{y}_n)\}_{n=1}^N,\qquad \hat{y}_n \in \{0,1\}^C,\qquad \sum_{i=1}^C \hat{y}_{n,i}=1.$$7, and AEVLP uses CLIP ViT-B/16; training runs for 10 epochs on CUB and NUS and 8 epochs on COCO and VOC (Tran et al., 28 Aug 2025).

5. Empirical behavior

The most direct evidence for GPR Loss is its performance across multiple pseudo-label sources. The paper evaluates random pseudo-labels, VLPL pseudo-labels, LL-Ct pseudo-labels, and DAMP pseudo-labels, then replaces the original loss with GPR. In all four cases, average mAP improves (Tran et al., 28 Aug 2025).

Pseudo-label source	Baseline average mAP	With GPR
Random	51.56	53.98
VLPL	58.53	59.13
LL-Ct	56.98	57.32
DAMP	58.31	59.90

These gains are reported as $$\hat{\mathcal{D}} = \{(x_n,\hat{y}_n)\}_{n=1}^N,\qquad \hat{y}_n \in \{0,1\}^C,\qquad \sum_{i=1}^C \hat{y}_{n,i}=1.$$8 for random pseudo-labels, $$\hat{\mathcal{D}} = \{(x_n,\hat{y}_n)\}_{n=1}^N,\qquad \hat{y}_n \in \{0,1\}^C,\qquad \sum_{i=1}^C \hat{y}_{n,i}=1.$$9 for VLPL, $\hat{y}_{n,i}=1$0 for LL-Ct, and $\hat{y}_{n,i}=1$1 for DAMP. This is the paper’s clearest empirical support for the claim that GPR is generalized across heterogeneous pseudo-label sources rather than tied to a single generator (Tran et al., 28 Aug 2025).

Within the full AEVLP framework, where DAMP supplies pseudo-labels and GPR provides the training objective, the reported mAP values are $\hat{y}_{n,i}=1$2 on VOC, $\hat{y}_{n,i}=1$3 on COCO, $\hat{y}_{n,i}=1$4 on NUS, and $\hat{y}_{n,i}=1$5 on CUB. The paper compares these with prior GR Loss values of $\hat{y}_{n,i}=1$6, $\hat{y}_{n,i}=1$7, $\hat{y}_{n,i}=1$8, and $\hat{y}_{n,i}=1$9, respectively. It also reports an ablation over GPR-specific components: positive pseudo-label reweighting $x_n \in \mathcal{X}$00, regularization term $x_n \in \mathcal{X}$01, and negative pseudo-label loss $x_n \in \mathcal{X}$02. Performance improves progressively as these components are added, and the final configuration gives the best results on all four datasets (Tran et al., 28 Aug 2025).

Further analysis varies the negative pseudo-label ratio $x_n \in \mathcal{X}$03. The reported best performance occurs around $x_n \in \mathcal{X}$04 to $x_n \in \mathcal{X}$05 pseudo-negatives across datasets, with relatively stable performance beyond that. A distribution plot described in the paper shows that AEVLP shifts positive-class probabilities closer to 1 while keeping negative distributions similar to GR Loss. The DAMP analysis reports that pseudo-label precision remains reasonably stable while accumulated recall increases across epochs, which the paper uses to motivate dynamic pseudo-label regeneration (Tran et al., 28 Aug 2025).

6. Relation to prior robust pseudo-label objectives and scope

GPR Loss directly extends GR Loss, the earlier SPML objective introduced in "Boosting Single Positive Multi-label Classification with Generalized Robust Loss" (Chen et al., 2024). GR Loss models each unlabeled class as a soft mixture of positive and negative supervision using a soft pseudo-label estimator $x_n \in \mathcal{X}$06, a reweighting term $x_n \in \mathcal{X}$07, and decoupled generalized losses $x_n \in \mathcal{X}$08. GPR preserves the GR branch for unresolved unlabeled entries but adds an explicit pseudo-label variable $x_n \in \mathcal{X}$09, together with distinct pseudo-negative and pseudo-positive branches (Tran et al., 28 Aug 2025).

The broader design principle of applying robust objectives specifically to pseudo-labeled data predates GPR. In "Semi-supervised Learning using Robust Loss" (Cui et al., 2022), the student is trained with standard CE on human-annotated labels and a robust loss—beta cross-entropy, generalized cross-entropy, or symmetric cross-entropy—on pseudo-labeled samples. That work is selective robustification in a teacher-student setting; GPR is a different formulation because it is designed for SPML and explicitly distinguishes observed positives, unresolved unlabeled entries, pseudo-negatives, and pseudo-positives.

A separate line of work, "Doubly Robust Self-Training" (Zhu et al., 2023), uses a correction-based pseudo-label objective rather than case-conditioned robust BCE-style branches. Its loss can be written as all-data pseudo-risk plus a labeled residual correction term. This suggests a different robust pseudo-label paradigm: debias pseudo-label learning by correction rather than by supervision-state decomposition. GPR does not use that estimator structure, but both approaches address the same problem of unsafe pseudo-label trust.

The limitations stated for GPR are specific. Its effectiveness still depends on pseudo-label quality; the weighting design introduces several hyperparameters; the exact schedules for some inherited parameters are not fully detailed in the main paper; the main-text theorem is qualitative; and the formulation is designed for SPML under the Assume Negative convention rather than for arbitrary partially labeled multi-label problems (Tran et al., 28 Aug 2025). Within that scope, however, GPR Loss is presented as a case-conditioned robust loss that learns from observed positives, unresolved unknowns, pseudo-negatives, and pseudo-positives in a single objective.