Intra-Class Contrastive Loss

• Intra-Class Contrastive Loss is a loss function that promotes intra-class compactness by reducing distances between same-class features while encouraging inter-class separation.
• It leverages various formulations—including pairwise Siamese, center-based, and proxy contrastive methods—to impose geometric constraints alongside traditional classification loss.
• Adaptive adaptations balance compactness with intra-class variability, enhancing robustness in applications ranging from speech emotion recognition to image classification.

Intra-class contrastive loss denotes a family of objectives that act directly on relations among samples, prototypes, centers, channels, or label embeddings associated with the same class. In the cited literature, the dominant formulation seeks intra-class compactness together with inter-class separability by pulling same-class representations together and pushing different-class representations apart; however, several later variants deliberately preserve or even enlarge intra-class variation when robustness, fine-grained discrimination, or richer soft labels are the primary objective (Lian et al., 2019, Qi et al., 2017, Duboudin et al., 2021, Yuan et al., 26 Sep 2025).

1. Definition and role in supervised representation learning

A recurring motivation is that standard classification supervision does not explicitly organize the geometry of learned features. In speech emotion recognition, cross-entropy with softmax is described as a supervision component that “does not explicitly encourage discriminative learning of features” (Lian et al., 2019). In image classification, softmax is described as a supervision signal that “only penalizes the classification loss,” leaving intra-class variations insufficiently constrained (Qi et al., 2017).

Intra-class contrastive loss addresses this gap by introducing an auxiliary or joint objective over similarities or distances. The canonical arrangement defines positive pairs as samples from the same class and negative pairs as samples from different classes. The optimization target is then to decrease positive-pair distance or increase positive-pair similarity, while enforcing a margin or a ratio against negatives. This basic design appears in pairwise Siamese objectives, center-based objectives, channel-based objectives, prototype-based objectives, and ordinal or long-tail adaptations (Lian et al., 2019, Song et al., 2020, Cai et al., 2023, Saleem et al., 2023).

A common implementation couples the contrastive term to a conventional classification loss. One explicit form is

$L=\lambda L_{\text{con}}+(1-\lambda)L_{\text{ce}},$

used in a Siamese CNN for speech emotion recognition (Lian et al., 2019). Another pattern is an auxiliary supervision signal added to softmax-based training, as in contrastive-center loss (Qi et al., 2017). In both cases, the contrastive term is not merely a classifier surrogate; it is a geometric constraint on the embedding space.

2. Canonical formulations

The most direct formulation is pairwise contrastive learning in a Siamese architecture. For a pair $(X_1,X_2)$, the shared encoder produces embeddings and the pair distance is

$D_W(X_1,X_2)=\|G_W(X_1)-G_W(X_2)\|.$

Two concrete variants were evaluated for speech emotion recognition. The cosine-based form is

$$L(m,(Y,X_1,X_2))= \begin{cases} 1-\cos(X_1,X_2) & Y=1,\ \max\bigl(0,\cos(X_1,X_2)-m\bigr) & Y=0, \end{cases}$$

with

$$\cos(X_1,X_2)=\frac{X_1\cdot X_2}{\|X_1\|_2\|X_2\|_2}.$$

The Euclidean-based form is

$$L(m,(Y,X_1,X_2))= \begin{cases} \max(0,m-\|X_1-X_2\|_2) & Y=0,\ \|X_1-X_2\|_2 & Y=1. \end{cases}$$

In both cases, positive pairs are explicitly contracted and negatives are repelled beyond a margin (Lian et al., 2019).

A second canonical family replaces instance pairs with class centers. In contrastive-center loss, each sample is compared with its corresponding class center and with all non-corresponding class centers: $$L_{ct-c}=\frac{1}{2}\sum_{i=1}^m \frac{\|x_i-c_{y_i}\|_2^2} {\left(\sum_{j=1,\, j\neq y_i}^{k}\|x_i-c_j\|_2^2\right)+\delta}.$$ The numerator reduces the distance to the correct class center, while the denominator increases the summed distance to incorrect centers, so intra-class compactness and inter-class separability are optimized simultaneously without explicit pair or triplet mining (Qi et al., 2017).

A more recent proxy formulation is Center Contrastive Loss, which maintains a class-wise center bank and compares the query to all class centers through a contrastive term,

$$\mathcal{L}^{\text{contrast}} = -\log \frac{\exp(c_y^T x/\tau)} {\exp(c_y^T x/\tau)+\sum_{j\ne y}\exp(c_j^T x/\tau)},$$

combined with a center constraint

$\mathcal{L}^{\text{center}}=\|x-c_y\|^2.$

The final objective is

$$\mathcal{L}=\mathcal{L}^{\text{contrast}}+\lambda \mathcal{L}^{\text{center}},$$

or, with additive margin, a CosFace-style variant using $(X_1,X_2)$0 and $(X_1,X_2)$1. This design replaces batch-limited sample mining with comparisons against globally maintained class proxies (Cai et al., 2023).

3. Structured variants: channels, ordinality, and label granularity

Several formulations relocate the intra-class constraint away from the final embedding. CC-Loss operates in the space of channel attention vectors rather than directly on feature embeddings. For a batch of $(X_1,X_2)$2 samples, channel attention vectors are stacked into $(X_1,X_2)$3, pairwise Euclidean distances are computed as

$(X_1,X_2)$4

and the batchwise intra- and inter-class terms are

$(X_1,X_2)$5

The complete loss is

$(X_1,X_2)$6

This formulation constrains “the specific relations between classes and channels” and uses matrix rewriting of $(X_1,X_2)$7 for efficient computation (Song et al., 2020).

Ordinal settings require a different treatment of negatives. SCOL augments supervised contrastive learning with a label-dependent distance metric,

$(X_1,X_2)$8

where

$(X_1,X_2)$9

Same-label samples are still pulled together, but negatives are penalized in proportion to ordinal distance. The paper states that this “remarkably enhanced inter-class separability and strengthened intra-class consistency among the AAC-24 genera” (Saleem et al., 2023).

The notion of “intra-class” can also be shifted from sample-level to label-level representations. In multi-label text classification, Intra-label Contrastive Loss (ICL) operates within each sample by pulling together the embeddings of labels that are present in the same text. It is optimized jointly with binary cross-entropy,

$D_W(X_1,X_2)=\|G_W(X_1)-G_W(X_2)\|.$0

This is not a class-clustering loss over sentence embeddings; it is an intra-sample constraint over label embeddings, introduced to improve “internal consistency for co-occurring labels” (Lin et al., 2022).

4. Sample selection, weighting, and imbalance-aware optimization

The effectiveness of intra-class contrastive learning depends strongly on how positives and negatives are selected and weighted. CACR modifies standard contrastive learning with a “doubly contrastive strategy.” Its contrastive attraction term weights more distant positives more heavily through a distribution $D_W(X_1,X_2)=\|G_W(X_1)-G_W(X_2)\|.$1, while contrastive repulsion weights closer negatives more heavily through $D_W(X_1,X_2)=\|G_W(X_1)-G_W(X_2)\|.$2. The practical effect is that hard positives and hard negatives contribute more than easy ones, rather than all positives and negatives being treated uniformly (Zheng et al., 2021).

This weighting is explicitly connected to robustness under imbalance. On exponential-imbalanced CIFAR-100, the reported performance drop is 12.57% for conventional CL and 9.24% for CACR with $D_W(X_1,X_2)=\|G_W(X_1)-G_W(X_2)\|.$3, which is presented as evidence that modeling intra-positive and intra-negative structure narrows the gap between sampled and target distributions (Zheng et al., 2021).

Long-tail settings introduce an additional issue: same-class compactness is often weakest for tail classes. Rebalanced Contrastive Learning addresses this by injecting class-frequency terms into supervised contrastive learning and by applying feature compression to underperforming tail classes. The rebalanced loss uses factors $D_W(X_1,X_2)=\|G_W(X_1)-G_W(X_2)\|.$4 and $D_W(X_1,X_2)=\|G_W(X_1)-G_W(X_2)\|.$5 derived from global class frequencies, while feature scaling $D_W(X_1,X_2)=\|G_W(X_1)-G_W(X_2)\|.$6 with $D_W(X_1,X_2)=\|G_W(X_1)-G_W(X_2)\|.$7 is used to draw tail samples closer together when validation accuracy is below 20% at half training epochs (Alvis et al., 2023).

Unsupervised person re-identification exposes a related problem: the model must decide whether to update memory using all samples, the hardest positive, or a safer intermediate sample. AdaInCV quantifies intra-class variation after clustering and introduces Adaptive Sample Mining and Adaptive Outlier Filter. AdaSaM selects harder positives when a cluster is tight and safer positives when it is loose, while AdaOF reuses valuable outliers as negative samples instead of discarding them. This replaces fixed update heuristics with an adaptive schedule based on measured intra-class difficulty (Liu et al., 2024).

5. Applications across domains

The literature applies intra-class contrastive objectives to image classification, speech emotion recognition, medical ordinal regression, multi-label text classification, person re-identification, speech embedding learning, long-tail recognition, class-incremental learning, image prediction, and knowledge distillation. The shared design pattern is that supervision is transferred from label identities to feature relations, but the object being contrasted differs by task: utterance embeddings, channel-attention vectors, class centers, label embeddings, prototypes, or aligned spatial patches (Lian et al., 2019, Song et al., 2020, Saleem et al., 2023, Lin et al., 2022, Andonian et al., 2021).

Setting	Intra-class mechanism	Reported result
Speech emotion recognition (Lian et al., 2019)	Siamese contrastive loss with cross-entropy	WA 62.19%, UWA 63.21%
AAC scoring (Saleem et al., 2023)	SCOL with ordinal-aware negatives	Pearson 89.04, Accuracy 85.27, F1 80.25
Multi-label text classification (Lin et al., 2022)	ICL over label embeddings within one sample	Macro F1 57.59, Micro F1 70.49, Jaccard 58.60
Speech embeddings (Zhang et al., 2023)	Contrastive loss plus ICC regularizer	EER 3.96% vs 4.39%
Unsupervised person Re-ID (Liu et al., 2024)	Adaptive intra-class variation contrastive learning	mAP 87.4% on Market-1501; 38.8% on MSMT17

Beyond these examples, long-tailed recognition uses prototype-based intra-branch contrastive loss over tail classes. In DB-LTR, adding metric loss alone yields 46.31% on CIFAR100-LT, metric plus intra-branch contrastive loss yields 47.48%, and combining metric, intra-branch, and inter-branch losses yields 48.88% (Chen et al., 2023). In class-incremental learning, BLCL combines a margin-based triplet-style constraint with Bayesian weighting between cross-entropy and contrastive terms and reports a Davies-Bouldin score of 0.81 and a Calinski-Harabasz index of 4,688 on CIFAR-10 cluster analysis (Raichur et al., 2024).

Paired structured prediction provides another extension. Contrastive Feature Loss for image prediction maximizes patchwise mutual information by treating corresponding spatial patches between prediction and ground truth as positives and other patches as negatives. This is an intra-class contrast at patch level rather than class-label level, and it is used as a drop-in replacement for $D_W(X_1,X_2)=\|G_W(X_1)-G_W(X_2)\|.$8-style critics (Andonian et al., 2021).

6. Theory, limitations, and alternative geometries

A common misconception is that intra-class contrastive loss is synonymous with maximal cluster collapse. The theoretical work on negatives-only supervised contrastive loss shows why that view is only partially correct. NSCL removes same-class samples from the denominator, and its global minimizers exhibit augmentation collapse, within-class collapse, and class centers that form a simplex equiangular tight frame. In this sense, one rigorous endpoint of supervised contrastive optimization is zero intra-class dispersion (Luthra et al., 4 Jun 2025).

At the same time, other work argues that this endpoint can be undesirable for fine-grained tasks. A direct comparison of contrastive and triplet loss reports that contrastive loss “tends to compact intra-class embeddings, which may obscure subtle semantic differences.” The measured mean intra-class variance is 0.031 for contrastive versus 0.074 for triplet on synthetic data, and 0.0030 versus 0.0059 on MNIST, with a paired $D_W(X_1,X_2)=\|G_W(X_1)-G_W(X_2)\|.$9-test yielding $$L(m,(Y,X_1,X_2))= \begin{cases} 1-\cos(X_1,X_2) & Y=1,\ \max\bigl(0,\cos(X_1,X_2)-m\bigr) & Y=0, \end{cases}$$0 (Zeng, 2 Oct 2025). This suggests that compactness is not an unconditional virtue; it trades off against detail retention.

Several methods therefore reverse or temper the standard objective. Reverse Contrastive Loss pushes same-class samples apart up to a safety condition,

$$L(m,(Y,X_1,X_2))= \begin{cases} 1-\cos(X_1,X_2) & Y=1,\ \max\bigl(0,\cos(X_1,X_2)-m\bigr) & Y=0, \end{cases}$$1

using the closest positive and closest negative in the batch. On the MNIST-MP benchmark, standard training yields test accuracy of about 26%, while RCL with $$L(m,(Y,X_1,X_2))= \begin{cases} 1-\cos(X_1,X_2) & Y=1,\ \max\bigl(0,\cos(X_1,X_2)-m\bigr) & Y=0, \end{cases}$$2 yields about 90%, even though validation accuracy decreases from about 99.8% to about 96% (Duboudin et al., 2021). Here, improved robustness is associated with larger intra-class spread rather than tighter clusters.

A related diversification argument appears in knowledge distillation. Teacher training with an intra-class contrastive loss is intended to enrich the intra-class information contained in soft labels, but the paper reports that intra-class loss causes instability in training and slows convergence. A margin condition based on $$L(m,(Y,X_1,X_2))= \begin{cases} 1-\cos(X_1,X_2) & Y=1,\ \max\bigl(0,\cos(X_1,X_2)-m\bigr) & Y=0, \end{cases}$$3 is therefore used to activate the intra-class term only when the teacher is sufficiently confident (Yuan et al., 26 Sep 2025). SimO offers a different compromise: it is anchor-free, optimizes both distance and orthogonality, and is reported to form class-specific, internally cohesive yet orthogonal neighborhoods that “balance class separation with intra-class variability” (Bouhsine et al., 2024).

This suggests that intra-class contrastive loss is best understood not as a single loss function, but as a design axis governing how much within-class variability should be removed, preserved, reweighted, or redistributed for a given task. Compactness is central in speaker verification, image classification, and many metric-learning settings, whereas preserved diversity becomes central in domain generalization, fine-grained retrieval, and distillation of richer soft labels (Zhang et al., 2023, Duboudin et al., 2021, Zeng, 2 Oct 2025, Yuan et al., 26 Sep 2025).