Quadruplet Loss-based Learning Approach

• Quadruplet loss-based learning is a metric learning framework that extends traditional triplet loss by incorporating additional negative constraints to improve class separation.
• Its methodology leverages advanced hard negative mining and dynamic margin strategies to optimize deep embeddings across various modalities and hierarchical labels.
• Empirical results in domains such as person re-identification, medical imaging, and federated learning demonstrate significant improvements in accuracy and robustness.

A quadruplet loss-based learning approach constitutes an extension of standard triplet-based metric learning protocols, strategically designed to enforce finer-grained relations among embedded samples in the context of supervised deep representation learning. Central to quadruplet loss designs is the simultaneous minimization of intra-class variance and maximization of inter-class variance, often across multiple dimensions (modalities, hierarchical labels, class prototypes, etc.), and the orchestration of multiple margin constraints. Across distinct application domains—person re-identification, cross-modal retrieval, federated learning, imbalanced classification, multi-output retrieval—the quadruplet loss paradigm offers improved generalization capability by leveraging constraints involving four samples, where the anchor–positive distance is penalized relative not only to a single negative but also with respect to additional negatives (often chosen across different classes, modalities, or patient-contexts) or to pairs from distinct semantic regions.

1. Mathematical Formulations and Principal Variants

Quadruplet losses are generally constructed from hinge-style constraints over four samples drawn according to the supervision regime. Let $x_i$ denote an anchor and $x_j$ a positive (often same label as $x_i$), with negatives $x_k$ and $x_l$ (from different identities or modalities). A canonical quadruplet loss for person re-identification is:

$$\begin{aligned} L_{\text{quad}} &= \sum_{s_i=s_j\neq s_k} \left[ g(x_i,x_j)^2 - g(x_i,x_k)^2 + \alpha_1 \right]_+ \ &\quad + \sum_{\substack{s_i=s_j,\, s_l\neq s_k,\, s_i\neq s_l,\, s_i\neq s_k}} \left[ g(x_i,x_j)^2 - g(x_l,x_k)^2 + \alpha_2 \right]_+, \end{aligned}$$

with margin parameters $\alpha_1 > \alpha_2$ and $g(x,y)$ representing the learned pairwise dissimilarity (Chen et al., 2017). This design generalizes the triplet loss by adding a secondary "push" in the space of distances under different anchors, thereby amplifying inter-class separation.

For complementary similarity learning tasks (Mane et al., 2019), quadruplet loss is decomposed into three explicit terms: similarity pull ($L_\text{sim}$), complementarity bounding ($L_\text{comp}$), and negative push ($L_\text{neg}$), each governed by separate margins $m_s, m_c, m_n$ and formulated on normalized item embeddings.

Distinct quadruplet structures are used in federated learning (with stochastic quadruplets pulling anchor-positive together and pushing anchor away from two negatives from different classes) (Goksu et al., 4 Sep 2025), metric learning for imbalanced data (Gui et al., 2021), patient-specific mining in medical imaging (Naranpanawa et al., 2023), ordinal retrieval of missing classes (Nazarovs et al., 2022), and robust face recognition under morphing attacks (Medvedev et al., 2024), each with domain-specific margin, sampling, and constraint definitions.

2. Quadruplet Construction, Sampling, and Mining Strategies

The efficacy of quadruplet loss is deeply tied to its sample selection. Adaptive online hard negative mining, such as Marg-OHNM (Chen et al., 2017), employs dynamically set margin values:

$$\alpha_1 = \mu_n - \mu_p,\quad \alpha_2 = 0.5(\mu_n - \mu_p),$$

where $\mu_p$ and $\mu_n$ are the batch means of positive and negative squared distances, selectively propagating gradients only for hard quadruplets exceeding the current model's margin threshold.

Hierarchical quadruplet selection (Karaman et al., 2019) mines the hardest negatives (minimally distant sample from a different coarse label) and relatively easy positives (same fine or coarse label, but farther in embedding space than the negative), using “outside-sphere” geometric constraints or ordering-based selection.

Uncertainty-based quadruplet selection leverages estimates of epistemic/aleatoric uncertainty from deep ensembles, choosing “similar” samples from classes with high uncertainty relative to the anchor, in addition to direct positives and negatives (Ott et al., 2024).

Dynamic margin design, used in patient-specific mining (Naranpanawa et al., 2023), computes per-patient margins $\alpha_x$ via k-means clustering over embeddings, stretching or shrinking loss sensitivity to individuated data geometry.

Cross-domain quadruplet mining (Cao et al., 2020) involves sampling negatives from different domains (NIR/VIS), selecting "hard" negatives via closest cosine similarity, and assembling quadruplets with both cross-domain and within-domain pulls and pushes.

3. Integration into Deep Network Architectures and Optimization

Quadruplet loss-based learning is usually realized within deep architecture frameworks with shared-weight Siamese or multi-branch subnetworks. Representative implementations include:

• An AlexNet-derived convolutional backbone, with a metric head emitting normalized similarity/dissimilarity scores over input image pairs (Chen et al., 2017).
• Siamese quadruplet towers for fashion recommendation, with Universal Sentence Encoder–based pre-embeddings followed by fully connected layers and $\ell_2$ normalization (Mane et al., 2019).
• Temporal models such as LSTM stacks and bidirectional RNNs, for time-series and imbalanced fault diagnosis (Gui et al., 2021, Nazarovs et al., 2022).
• CNNs for metric learning in visual domains, incorporating decorrelation learning via a decorrelation layer and shared projection matrices across modalities (Cao et al., 2020).
• Meta-learning architectures with external memory and GRU-based controllers for margin estimation in zero-shot sketch-based image retrieval (Liu et al., 2024).

Optimization combines the quadruplet loss, sometimes with auxiliary softmax/cross-entropy classification objectives, using standard optimizers (Adam, SGD). Training procedures often employ data augmentation and adaptive sampling, and may include additional regularizers (global mean/variance penalties (Karaman et al., 2019), prototype whitening/re-coloring (Palit et al., 2022)).

4. Comparative Impact Versus Triplet and Contrastive Losses

Quadruplet loss amplifies metric learning signal beyond the scope of triplet and pairwise losses by introducing multiple negative constraints, refining the embedding’s separation power. Where contrastive losses independently push negatives and pull positives, they lack relative ordering imposed by hard negatives. Triplet loss introduces relative comparison but remains limited to a single negative per anchor. Quadruplet loss, with dual or higher-order negative directions (multi-modality, cross-identity, prototype-smoothing, hierarchical semantic criteria), produces superior cluster geometry and minimizes class overlap (Chen et al., 2017, Goksu et al., 4 Sep 2025, Karaman et al., 2019, Palit et al., 2022, Naranpanawa et al., 2023).

Quantitative results consistently show higher retrieval, clustering, and classification sensitivity across diverse metrics and datasets; e.g., the quadruplet network improves rank-1 accuracy for person ReID on CUHK03 by 2.7 percentage points over triplet (Chen et al., 2017), boosts recall@1 on fine-grained image retrieval to 66% compared to 61% for triplet+global loss (Karaman et al., 2019), and nearly doubles missing-class accuracy in ordinal time-series classification (Nazarovs et al., 2022).

5. Advanced Margin Schemes and Meta-Learning Extensions

Several works implement adaptive, dynamic, or meta-learned margin strategies:

• Data-driven margins from batch-level statistics, e.g., $\alpha_1 = \mu_n – \mu_p$ (Chen et al., 2017).
• Per-patient dynamic margins computed via k-means centroid separation (Naranpanawa et al., 2023).
• Margin meta-learning using memory-augmented networks (RAMLN), which read from external memory to optimize per-batch or per-class loss margins (Liu et al., 2024).
• Weight and margin parameters in face-morphing security, where multi-term hinge loss balances anchor/positive/negative/morph distances with learnable weights (Medvedev et al., 2024).

These approaches aim to maintain optimal separation where class geometry is dynamic or non-uniform, or where class-wise adaptation is required (cross-patient, cross-modality, incremental-class scenarios).

6. Multimodal, Hierarchical, and Semantic Quadruplet Extensions

Quadruplet losses have been customized for semantically coherent embedding in multi-output problems (Proença et al., 2020), hierarchical label structures (Karaman et al., 2019), multimodal retrieval (e.g., photo-sketch (Liu et al., 2024)), and cross-domain face recognition (Cao et al., 2020). Semantic disagreement is quantified via $l_0$-norm between label vectors, yielding feature space geometry directly mirroring semantic overlap.

In zero-shot and incremental learning contexts (few-shot class incremental (Palit et al., 2022), missing class retrieval (Nazarovs et al., 2022)), quadruplet constraints among prototypes or class centers with decorrelation and whitening regularizers have demonstrated strong resistance to catastrophic forgetting and improved accuracy on previously unseen classes.

7. Application Domains and Empirical Gains

Quadruplet-loss–based approaches are validated across diverse domains:

Domain	Principal Design Features	Noted Quantitative Gains
Person re-identification	Auxiliary margin, Marg-OHNM, normalized metric head	+2.7 pp rank-1 accuracy on CUHK03
Fashion recommendation	Similarity/complementarity negatives, tight margin design	+30 pp ranking acc over triplet baselines
Federated learning (FedQuad)	Local quadruplet construction, global averaging	3–6 pp accuracy lift under non-IID splits
Few-shot/interference learning	Uncertainty-based quadruplet mining ensemble	97.66% accuracy, +0.07 F₂-score over triplet
Medical imaging (DMT-Quadruplet)	Tiered quadruplet, patient-specific dynamic margin	+54% sensitivity to rare class (UD)
Ordinal time-series	Log-ratio ordinal constraint, missing-class retrieval	Doubled accuracy on missing-class detection
Robust face rec. (morphing)	Morph-augmented quadruplet, multi-term margin	Competitive MMPMR, robust against attacks
Fine-grained/image retrieval	Hard-negative quadruplet mining, hierarchical labels	+4–6 pp Recall@1 over random and triplet
Multi-label semantic embedding	Semantic ordering via $l_0$ norm, quadruplet hinge	+2–4 pp mAP in LFW/MegaFace over triplet

These results consistently confirm that quadruplet loss frameworks deliver enhanced generalization, retrieval, and robustness, especially in contexts characterized by severe class imbalance, cross-domain structure, hierarchical semantics, or rapid class turnover.

---

Quadruplet loss-based learning approaches have fundamentally broadened the representational power of deep embedding learning, by encoding multi-way margin constraints and advanced hard sample mining. Empirical studies in surveillance, fashion, medical imaging, federated and few-shot learning robustly support their superiority over traditional metric learning frameworks, with extensions to cross-modal, multi-label, and dynamically adaptive regimes driving ongoing research.